CN118677787A - A model training method and device - Google Patents

Abstract

The embodiments of the present application provide a model training method and device for quickly improving the accuracy of the model so that the network can quickly converge to the desired target, thereby improving the intelligence of network control. The first network element uses the network behavior (such as the first instruction) as an input parameter to train the network prediction model, which can quickly improve the prediction accuracy of the network prediction model under the network behavior, so that the network can quickly converge to the desired target. The second network element trains the network control model based on the behavioral effect of the network behavior (such as the change in at least one network indicator corresponding to the first instruction) and the change in the target network indicator that can achieve the network operation target, which can quickly improve the control accuracy of the network control model so that the network can quickly converge to the desired network operation target.

Description

Model training method and device

Technical Field

The present application relates to the field of communications technologies, and in particular, to a model training method and apparatus.

Background

The network data analysis is a service for collecting relevant network data from network function network elements, application function network elements or a network management system according to standard specifications, performing association analysis on the collected data with various dimensions by using machine learning and big data mining technologies, training and fitting a model, and outputting analysis or prediction according to the model. The models can be further divided into predictive models and control models, among others. The prediction model can train a fitting model by utilizing historical or current network data, so that the prediction model predicts the running state of the network and is an important component of the network data analysis network element; the control model is used to describe the impact that a particular network action, if taken, will have on future network operations, to simulate the consequences of that particular network action, and to decide whether this network action should be taken, is an important component of the network control network element. Whether a predictive model or a control model, the model can be obtained by performing model training through a large amount of network data by using a machine learning technology.

However, in a specific implementation, it is difficult to train an accurate prediction model from the beginning due to limited network data collected or traffic differences of various areas; on the other hand, it is difficult to accurately model the network element behavior and how the network operates at first, and to accurately evaluate the impact of various network behaviors on the network operation, so that it is difficult to decide what network behavior should be taken. Therefore, the prior art relies on an accurate model from the beginning, but the reality is that the accurate model is difficult to be provided initially, so that the network control intellectualization is difficult to realize.

Disclosure of Invention

The embodiment of the application provides a model training method and device, which are used for rapidly improving the precision of a model, so that a rapid network is converged to a desired target, and the intelligence of network control is further improved.

In a first aspect, a model training method is provided, including: the first network element determines a second network index according to a first network prediction model, wherein the input parameters of the first network prediction model comprise the first network index, the first network index comprises an actual network index of the network at a first moment and/or an actual network index before the first moment, the second network index is a network index of the network output by the first network prediction model according to the first network index at a second moment, and the second moment is later than the first moment; the first network element determines a first prediction deviation, wherein the first prediction deviation is the deviation of a second network index relative to a third network index, the third network index is an actual network index of the network at a second moment, and the network executes a first instruction between the first moment and the second moment, and the first instruction is used for adjusting network parameters of the network; the first network element adjusts model parameters of the first network prediction model according to the first prediction deviation and the first instruction to obtain a second network prediction model; when the input parameters of the second network prediction model comprise a first network index and a first instruction, the deviation of the network index of the network output by the second network prediction model at the second moment relative to the third network index is smaller than the first prediction deviation.

In the scheme, the network behavior (such as the first instruction) is used as an input parameter to train the network prediction model, so that the prediction accuracy of the network prediction model under the network behavior can be rapidly improved, and the network can be rapidly converged to a desired target. Thus, the network can be allowed to realize network intellectualization by using a low-precision network prediction model at first, and the intellectualization of network control can be improved.

In one possible design, the first network element may cluster the instructions according to the first instruction, or the first network element clusters the instructions according to the first network index and the second network index to obtain a first class of instructions; wherein the first instruction belongs to the first class of instructions. And the first network element can adjust model parameters of the first network prediction model according to the first prediction deviation and the first type of instructions to obtain a second network prediction model. Optionally, the network index change amount corresponding to each instruction in the first type of instruction is within a preset range.

Therefore, the prediction accuracy of the network prediction model under the same type of network behavior (such as a first type of instruction) can be rapidly improved, and the model training efficiency can be improved.

In one possible design, the first network element may send the first network indicator and/or the second network indicator to the second network element; the first network element may receive a first instruction from the second network element, wherein the first instruction is output by a network control model running on the second network element, and input parameters of the network control model include a first network indicator and/or a second network indicator, and further include a network running target.

Thus, the first network element can obtain the first instruction from the second network element, and the reliability of the scheme is improved. In addition, the linkage adjustment of the network prediction model and the network control model is facilitated, so that the cooperation of prediction and control is more accurate.

In one possible design, the network operation target includes a network operation target corresponding to the target user, and the first instruction includes a network adjustment parameter corresponding to the target user.

Thus, the overall operation performance of the network and the service experience of the target user can be considered.

In one possible design, the first network element may send a second network index to the third network element, where the second network index is used for outputting, by the user prediction model, a network index corresponding to the target user at the second moment.

Therefore, the network index of the target user can be predicted, and the network service quality of the target user is better ensured.

In one possible design, the input parameters of the first network prediction model further include a first instruction, and the second network index is a network index of the network output by the first network prediction model at the second moment according to the first network index and the first instruction. In other words, the network prediction models before and after adjustment each have the network behavior as an input parameter.

Therefore, the network prediction model can be trained for multiple times, so that the prediction deviation of the network prediction model under the condition that the network executes the first instruction is smaller and smaller, and the accuracy of the network prediction model is further improved.

In a second aspect, a model training method is provided, including: the second network element determines a first instruction according to the first network control model; the input parameters of the first network control model comprise a network operation target and at least one network index, the first instruction is an instruction which is output by the first network control model and is used for speculatively enabling the at least one network index to reach the network operation target, and the first instruction is used for adjusting the network parameters of a network; the second network element determines a first control deviation, wherein the first control deviation is the deviation of the change amount of at least one network index corresponding to the first instruction relative to the change amount of a target network index capable of achieving a network operation target; the second network element adjusts the model parameters of the first network control model according to the first control deviation to obtain a second network control model; when the input parameters of the second network control model comprise a network operation target and at least one network index, the deviation of the change amount of at least one network index corresponding to the second instruction output by the second network control model relative to the change amount of the target network index is lower than the first control deviation.

In the above scheme, the second network element determines the control deviation of the network control model based on the behavior effect of the network behavior (such as the change amount of at least one network index corresponding to the first instruction) and the target network index change amount capable of achieving the network operation target, and trains the network control model based on the control deviation, so that the control precision of the network control model can be rapidly improved, and the network can be rapidly converged to the desired network operation target.

In one possible design, at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of at least one network index is the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index.

In the design, the change amount of the actual network index generated by the first instruction is used as the behavior effect of the first instruction, and the difference value between the first network index and the network operation target before the first instruction is executed is used as the target network index change amount, so that the implementation is simple and the implementation is easy.

In one possible design, at least one network index is a second network index, where the second network index is a network index of the network at a second moment, and the second moment is later than the first moment, where the network prediction model outputs the network according to an actual network index of the network at the first moment and/or an actual network index before the first moment; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the second network index.

In the design, the change amount of the network prediction index generated by the first instruction is used as the behavior effect of the first instruction, and the difference value between the network prediction index and the network operation target before the first instruction is executed is used as the target network index change amount, so that the implementation is simple and the implementation is easy.

In one possible design, the at least one network indicator includes a first network indicator and a second network indicator, wherein the first network indicator is an actual network indicator of the network at a first time and/or an actual network indicator before the first time, and the second network indicator is a network indicator of the network output by the network prediction model according to the first network indicator at a second time, and the second time is later than the first time; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction or the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index.

In the design, the network prediction index generated by the first instruction or the change amount of the actual network index is used as the behavior effect of the first instruction, and the difference value between the actual network index and the network operation target before the first instruction is executed is used as the change amount of the target network index, so that the implementation mode is flexible, simple and easy to implement.

In one possible design, initial values of model parameters of the first network control model are determined based on network intent. The network intention is, for example, an intention of the network operator to the network entered by the network operator. In other words, the first network control model may be a low-precision network control model.

Therefore, network intellectualization based on a low-precision network control model can be realized, and the intellectualization of network control can be improved.

In one possible design, the second network element may send a first instruction to the first network element, the first instruction being for adjusting model parameters of a network prediction model running on the first network element.

Therefore, the linkage adjustment of the network prediction model and the network control model is facilitated, and the cooperation of prediction and control is more accurate.

In one possible design, the network operation target includes a network operation target corresponding to the target user, the first instruction includes a network adjustment parameter corresponding to the target user, and the input parameter of the first network control model further includes a network adjustment parameter corresponding to the target user. The second network element may further obtain a network adjustment parameter corresponding to the target user from the fourth network element, where the network adjustment parameter is output by a user control model running on the fourth network element.

Therefore, the network index of the target user can be adjusted, and the network service quality of the target user is better ensured.

In a third aspect, there is provided a communication device comprising means or units or technical means for implementing the method described in the first aspect or any one of the possible designs of the first aspect.

Illustratively, the communication device may include:

The determining module is used for determining a second network index according to the first network prediction model, wherein the input parameters of the first network prediction model comprise first network indexes, the first network indexes comprise actual network indexes of the network at a first moment and/or actual network indexes before the first moment, the second network indexes are network indexes of the network at a second moment, which are output by the first network prediction model according to the first network indexes, and the second moment is later than the first moment;

The training module is used for determining a first prediction deviation, wherein the first prediction deviation is the deviation of a second network index relative to a third network index, the third network index is an actual network index of the network at a second moment, and the network executes a first instruction between the first moment and the second moment, and the first instruction is used for adjusting network parameters of the network; according to the first prediction deviation and the first instruction, model parameters of the first network prediction model are adjusted, and a second network prediction model is obtained; when the input parameters of the second network prediction model comprise a first network index and a first instruction, the deviation of the network index of the network output by the second network prediction model at the second moment relative to the third network index is smaller than the first prediction deviation.

In a fourth aspect, there is provided a communication device comprising means or units or technical means for implementing the method described in the second aspect or any one of the possible designs of the second aspect.

Illustratively, the communication device may include:

The determining module is used for determining a first instruction according to the first network control model; the input parameters of the first network control model comprise a network operation target and at least one network index, the first instruction is an instruction which is output by the first network control model and is used for speculatively enabling the at least one network index to reach the network operation target, and the first instruction is used for adjusting the network parameters of a network;

the training module is used for determining a first control deviation, wherein the first control deviation is the deviation of the change amount of at least one network index corresponding to the first instruction relative to the change amount of a target network index capable of achieving a network operation target; adjusting model parameters of the first network control model according to the first control deviation to obtain a second network control model; when the input parameters of the second network control model comprise a network operation target and at least one network index, the deviation of the change amount of at least one network index corresponding to the second instruction output by the second network control model relative to the change amount of the target network index is lower than the first control deviation.

In a fifth aspect, there is provided a communication device comprising a processor and interface circuitry for receiving signals from or transmitting signals from a communication device other than the processor for implementing the method as described in the first aspect or any one of the possible designs of the first aspect or the second aspect or any one of the possible designs of the second aspect, by logic circuitry or executing code instructions.

In a sixth aspect, an embodiment of the present application provides a communications device comprising a processor coupled to a memory, the processor being configured to invoke a program stored in the memory to perform a method as described in the first aspect or any of the possible designs of the first aspect or the second aspect or any of the possible designs of the second aspect. The memory may be located within the device or may be located external to the device. And the processor may be one or more.

In a seventh aspect, there is provided a computer readable storage medium having stored therein a computer program or instructions which, when executed by a communication device, implement a method as described in the first aspect or any one of the possible designs of the first aspect or the second aspect or any one of the possible designs of the second aspect.

In an eighth aspect, embodiments of the present application also provide a computer program product comprising a computer program or instructions which, when executed by a communication device, cause the method as described in the first aspect or any one of the possible designs of the first aspect or the second aspect or any one of the possible designs of the second aspect to be performed.

A ninth aspect provides a communication system comprising a first network element for performing the method as described in the first aspect or any one of the possible designs of the first aspect, and a second network element for performing the method as described in the second aspect or any one of the possible designs of the second aspect.

The advantages of the above third to ninth aspects and any possible designs thereof can be seen from the description of the advantages of the first to second aspects and any possible designs thereof, and are not specifically developed.

Drawings

Fig. 1 is a schematic architecture diagram of a communication system to which an embodiment of the present application is applied;

FIG. 2 is a schematic diagram of a deep learning model;

FIG. 3 is a schematic diagram of a network intelligent control scheme;

fig. 4 is a schematic diagram of a network control system according to an embodiment of the present application;

FIG. 5 is a flowchart of a model training method according to an embodiment of the present application;

FIG. 6 is a flow chart of a second round of training;

FIG. 7 is a flowchart of another model training method according to an embodiment of the present application;

FIGS. 8A and 8B are exemplary diagrams of another model training method provided by an embodiment of the present application;

FIGS. 9A and 9B are exemplary diagrams of another model training method provided by an embodiment of the present application;

FIGS. 10A and 10B are exemplary diagrams of another model training method provided by an embodiment of the present application;

Fig. 11 is a schematic structural diagram of a communication device according to an embodiment of the present application;

fig. 12 is a schematic structural diagram of another communication device according to an embodiment of the present application.

Detailed Description

The embodiment of the application can be applied to a mobile communication network (such as a 5th generation (5th generation,5G) mobile communication network or a 6 th generation (5th generation,6G) mobile communication network, etc.), a fixed communication network, a satellite communication network, etc.

Taking a 5G mobile communication network as an example, fig. 1 is a schematic architecture diagram of a communication system 1000 to which an embodiment of the present application is applied. As shown in fig. 1, the communication system 1000 includes a core network that includes one or more of the following entities:

Access and mobility management function network elements: the method is mainly used for the attachment, mobility management and tracking area updating flow of the terminal in the mobile network. The access and mobility management function network element processes non-access stratum (non access stratum, NAS) messages, performs registration management, connection management, and reachability management, allocates tracking area lists (TRACK AREA LIST, TAlist), and mobility management, etc., and transparently routes session management (session management, SM) messages to the session management network element. In the 5th generation (5th generation,5G) communication system, the access and mobility management function network element may be an access and mobility management function (ACCESS AND mobility management function, AMF).

Session management network element: the method is mainly used for session management in the mobile network, such as session establishment, modification and release. Specific functions are for example assigning an internet protocol (internet protocol, IP) address to the terminal, selecting a user plane network element providing a message forwarding function, etc. In a 5G communication system, the session management network element may be a session management function (session management function, SMF).

Policy control network element: including subscriber subscription data management functions, policy control functions, charging policy control functions, quality of service (quality of service, qoS) control, etc. In a 5G communication system, the policy control network element may be a policy control function (policy control function, PCF). It should be noted that the PCF in an actual network may also be hierarchically or functionally divided into multiple entities, e.g. a global PCF and a PCF within a slice, or a session management PCF (MANAGEMENT PCF, SM-PCF) and an access management PCF (ACCESS MANAGEMENT PCF, AM-PCF).

Network slice selection network element: the method is mainly used for selecting the proper network slice for the service of the terminal. In a 5G communication system, the network slice selection network element may be a network slice selection function (network slice selection function, NSSF) network element.

Unified data management network element: is responsible for managing subscription information of the terminal. In a 5G communication system, the unified data management network element may be unified data management (unified DATA MANAGEMENT, UDM).

Data analysis network element: the data analysis network element collects network data from various Network Functions (NFs), such as AMF, SMF, PCF, etc. The data analysis network element may collect network data indirectly from the application function (application function, AF) or directly from the AF through a network development function (network exposure function, NEF); the data analysis network element may also collect network data from an operations, administration, AND MAINTENANCE, OAM, system. The data analysis network element may use a machine learning technique to perform training and fitting on the collected network data to form a network prediction model, and then output a prediction result according to the network prediction model. The prediction result provided by the network data analysis network element can be used for adjusting the network and optimizing the service experience, and the network intellectualization is promoted. In a 5G communication system, the data analysis network element may be a network data analysis function (network DATA ANALYTICS function, NWDAF), or a management data analysis system (MANAGEMENT DATA ANALYTICS SYSTEM, MDAS).

User plane network element: the method is mainly responsible for processing the user message, such as forwarding, charging, legal monitoring and the like. The user plane network element may also be referred to as a protocol data unit (protocol data unit, PDU) session anchor (PDU session anchor, PSA). In a 5G communication system, the user plane network element may be a user plane function (user plane function, UPF). UPF may communicate directly with NWDAF via a similarly serviced interface, or may communicate with NWDAF via other approaches, such as via a proprietary or internal interface between SMF or and NWDAF.

Application function network element: mainly supporting interaction with the third generation partnership project ((3rd generation partnership project,3GPP) core network to provide services, such as affecting data routing decisions, policy control functions, or providing some service to a third party on the network side, in a 5G communication system, the application function network element may be an AF.

Network element with open function: the method is mainly used for supporting the opening of the capability and the event, such as safely opening the service and the capability provided by the 3GPP network function to the outside. In a 5G communication system, the network element may also be a network open function (network exposure function, NEF).

Network storage function network element: the method is mainly used for storing the network function entity and the description information of the service provided by the network function entity, supporting service discovery, network element entity discovery and the like. In a 5G communication system, the network storage function network element may be a network storage function (network repository function, NRF).

Operation management maintenance network element: the method is mainly used for managing resource configuration, performance statistics, fault alarms and the like of the network equipment. In the 5G communication system, the operation administration and maintenance network element may be OAM or the like.

In addition to the entities involved in the core network described above, the communication system 1000 may include the following devices or network elements:

And (3) a terminal: is a device with wireless receiving and transmitting function. A terminal may also be referred to as a terminal device, user Equipment (UE), mobile station, mobile terminal, etc. The terminal may be widely applied to various scenes, for example, device-to-device (D2D), vehicle-to-device (vehicle to everything, V2X) communication, machine-type communication (MTC), internet of things (internet of things, IOT), virtual reality, augmented reality, industrial control, autopilot, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, and the like. The terminal can be a mobile phone, a tablet personal computer, a computer with a wireless receiving and transmitting function, a wearable device, a vehicle, an unmanned aerial vehicle, a helicopter, an airplane, a ship, a robot, a mechanical arm, intelligent household equipment and the like. The embodiment of the application does not limit the specific technology and the specific equipment form adopted by the terminal. For convenience of description, a terminal will be described below as an example.

Access Network (AN) device: the radio side access used for the terminal may be a base station (base station), an evolved NodeB (eNodeB), a transmission and reception point (transmission reception point, TRP), a next generation NodeB (gNB) in a fifth generation (5th generation,5G) mobile communication system, a next generation base station in a sixth generation (6th generation,6G) mobile communication system, a base station in a future mobile communication system, or an access node in a wireless fidelity (WIRELESS FIDELITY, wiFi) system; the present application may also be a module or unit that performs a function of a base station part, for example, a Central Unit (CU) or a Distributed Unit (DU). The CU here performs the functions of radio resource control (radio resource control, RRC) protocol and packet data convergence layer protocol (PACKET DATA convergence protocol, PDCP) of the base station, and may also perform the functions of service data adaptation protocol (SERVICE DATA adaptation protocol, SDAP); the DU performs functions of a radio link control (radio link control, RLC) layer and a medium access control (medium access control, MAC) layer of the base station, and may also perform functions of a partial Physical (PHY) layer or all physical layers, and for a detailed description of each of the above protocol layers, reference may be made to related technical specifications of 3 GPP. The access network device may be a macro base station, a micro base station, an indoor station, a relay node, a donor node, or the like. The specific technology and specific device configuration adopted by the access network device in the embodiment of the application are not limited. For convenience of description, a base station will be described below as an example of an access network device.

The Data Network (DN) may be a service network providing data traffic services for users. For example, the DN may be an IP multimedia service (IP multi-MEDIA SERVICE) network or the Internet (Internet), etc. Wherein the terminal device may establish a protocol data unit (protocol data unit, PDU) session from the terminal device to the DN to access the DN.

It should be understood that the network elements illustrated in fig. 1 are merely examples and are not limiting, and other network elements may be included in the network architecture in practical applications, which is not limited by the present application.

For convenience of description, each network element may be expressed by its corresponding english abbreviation, for example, by "NWDAF" for a data analysis network element, by "SMF" for a session management function network element, and so on.

In order to realize intelligent control of the network, the network control network element is required to take corresponding actions according to the output of the network data analysis network element and optimize the network operation besides the data analysis network element to output the prediction result.

For example, fig. 3 is a schematic diagram of a network intelligent control scheme. The central network automation function network element (the central network automation function network element is an independent network control network element) can output recommended actions which each network function should take according to the network control model and the forecast of NWDAF network elements, so as to achieve the network automation goal. The network element (e.g. AMF, SMF, PCF, etc.) processing the service obtains the recommended value of outputting certain data, and adjusts the network operation according to the recommended value, finally achieving the expected effect. For example, after obtaining the desired quality of experience level of the application and the predicted future quality of experience level of the application NWDAF in the specified location area, the central network automation function network element determines a recommended value of the network operation information in the specified location area (i.e., the network KPI that should be reached) based on a model of an association between the application information (i.e., the quality of experience of the application) and the network operation information (e.g., key performance indicators (Key Performance Indicator, KPI) of the network). In this association model, the argument may be a factor such as a UE location, an application location, a bit rate, a packet delay, a number of transmission and retransmission messages, etc., and the argument may be an application experience quality, such as an average subjective evaluation (Mean Opinion Score, MOS). The central network automation function network element can predict NWDAF network elements and determine recommended values such as the application position, bit rate, packet delay, retransmission rate and the like when the recommended MOS value reaches the expected value. These factors affecting the quality of service experience are controlled by a plurality of service processing network elements (e.g., PCF, SMF, UPF, etc.), which perform network adjustment according to the recommended actions of the central network automation function network element, so as to achieve the network reaching the desired target value.

Whether it be a network analysis network element or a network control network element, specific functions are implemented according to the model. The model is the result of abstract descriptions of network operation rules and network behavior, and can be used to predict the impact of network operation and specific network behavior on network operation. For example, the network prediction model may use historical or current network data to train a fitting model, so as to predict the running state of the network, which is an important component of the network data analysis network element; the network control model is used to describe the impact that a particular network action, if taken, will have on future network operations, to simulate the consequences of that particular network action, and to decide whether this network action should be taken, is an important component of the network control network element.

Whether a network prediction model or a network control model can be obtained by a method of training a model by using a large amount of network data by using a machine learning technology. In this process, the network elements of the network prediction model or the network control model operate in a Training mode (Training mode). Taking training the deep learning Model (DEEP LEARNING Model) shown in fig. 2 as an example, running in training mode, the training process includes:

First, the network element responsible for the network prediction model collects earlier network operation data (e.g., network throughput rate before one week) as an argument (INDEPENDENT VARIABLES) and later or current network operation data (e.g., network throughput rate before one day or current) as an argument (DEPENDENT VARIABLES); or historical data of one part of the dimension (such as data rate, time delay, packet loss rate, etc. of a part of the users historically) is taken as an independent variable (INDEPENDENT VARIABLES), and historical data of another part of the dimension (such as business experience score of a part of the users, etc.) is taken as an independent variable (DEPENDENT VARIABLES); or network operational data before taking network action (e.g. the service experience score before modifying the maximum data rate allowed by the service flow of a part of users, the maximum data rate allowed by the original service flow, the maximum data rate allowed by the increased or decreased service flow) is taken as an independent variable, and network operational data after taking network action (e.g. the service experience score after increasing or decreasing the maximum data rate allowed by the service flow) is taken as a dependent variable.

Then, the network element responsible for the network prediction model or the network element responsible for the network control model trains the deep learning model by using the relevant data after analysis and mining, and the model parameters of the deep learning model are obtained according to the collected data, namely, the optimal weight parameters (WEIGHT PARAMETERS) of each neuron of each layer in the deep learning model are obtained, so that after the self-variable values of the training data are input according to the weight parameters, the model can output the values of the dependent variables with the minimum deviation, and the training is finished.

After training, the model may be run in inference mode (INFERENCE MODE). For example, the network prediction model can calculate future network operation data (e.g., network throughput rate after 1 hour) as a dependent variable based on historical or current network operation data (e.g., network throughput rate before a day or current) as an independent variable; or from data of one dimension as an argument (e.g., data rate, delay, packet loss rate, etc. of the current part of users), data of another dimension as an argument (e.g., current or future business experience scores of the current part of users, etc.). For another example, the network control model can calculate network operation data (e.g., a traffic experience score after increasing or decreasing a maximum data rate allowed by a traffic flow) after the network operation is expected to be taken according to the current network operation state as an argument and the network operation to be taken (e.g., a traffic experience score of a current partial user, a maximum data rate allowed by a current traffic flow, a maximum data rate allowed by a traffic flow to be increased or decreased), and if the network operation data can reach a target value, output the network operation to be taken; or the network control model can calculate the impact on the network operation data (e.g., the change in the traffic experience score after increasing or decreasing the maximum data rate allowed by the traffic flow) after the network action is expected to be taken based on the network action to be taken (e.g., the maximum data rate allowed by the traffic flow) as an argument, and if the impact on the network operation data by the current network operation data overlap can reach the target value, the network action to be taken is output.

In the above-described network intelligent control scheme, two accurate models are required to achieve the desired target values. The first model is an accurate network prediction model, namely, the network prediction model is required to accurately predict the future running condition of the network according to the running state of the network and the business input; the second model is an accurate network control model of each service processing network element, namely, the network control model is required to accurately output the influence of the network element on the network operation according to the recommended action. In a specific reality, however, it is difficult to train an accurate network prediction model from the beginning because of limited collected network data or business differences of various areas; on the other hand, it is difficult to accurately model the network element behavior and how the network operates at first, and it is difficult to accurately evaluate the impact of various network actions on its network operation. Therefore, the above solution relies on an accurate model from the beginning, but it is difficult to have an accurate model in the beginning, so that an intelligent network control system is difficult to realize.

In view of this, the technical scheme of the embodiment of the application is provided. The embodiment of the application allows the network to initially use a low-precision model (comprising a network prediction model and/or a network control model) to control the operation of the network, and takes the behavior (such as an execution instruction) executed by the network as the input of the model to train the model, so that the precision of the model can be quickly improved, the network can be quickly converged to a desired target, and the intelligence of network control is further improved.

Referring to fig. 4, a schematic diagram of a network control system according to an embodiment of the present application includes a network, a first network element, and a second network element.

The network may be any physical communication network (e.g., a 5G network, etc.), where the network is formed by one or more network functions, and the network functions receive service connection and data messages initiated by a user, process the service state and the data messages, generate updated service state and processed data messages, and send the data messages to the user. In the application embodiment, the network may be an entire communication network deployed by an operator, or may be a part of a communication network, for example, a part of a network in a certain area, or a network slice of a certain tenant or a certain service, or a part of a network involved in a certain communication service. From the network as a whole, it has traffic inputs and outputs at the current time, and the current traffic state. After service processing, the method has the processed service input and output and updated service state. The service input and output of the network system can measure certain network indexes, such as network load percentage, average data receiving and transmitting rate of each user in the network, time delay of service message in network transmission and the like; in addition, the network status may also be measured and controlled, such as whether the network is in a congested state, the network congestion level, the status of the user initiated connection, and so on. Wherein the measurements may be periodically-spaced metrics and status acquisitions, such as network metrics and network status once per second; the measurement may also be event triggered, such as an event notification after network congestion has been entered, or an event notification that the network throughput rate has reached a certain value. The measured network metrics and/or network status may be provided to the first network element and/or the second network element for use.

The first network element is used for outputting a network prediction index. For example, the first network element obtains an actual network indicator and/or network status (referred to herein as a network measurement indicator), predicts a network indicator and/or network status (referred to herein as a network prediction indicator) of the network operating at a certain moment in the future, such as a network load percentage prediction after 1 minute, a network congestion level prediction, etc., in accordance with a network prediction model in combination with the actual network indicator and/or network status.

In specific implementation, the first network element may be a module of a network element for processing services in the network or a network control network element (Network Control Function), may also be an independent data analysis network element (such as NWDAF, etc.), may also be a network element for processing services in the network (such as AMF, SMF, PCF, etc.) itself or a module thereof, etc., which is not limited in this aspect of the present application. For ease of description, the following will mainly take the case that the first network element is a separate data analysis network element (e.g. NWDAF).

The second network element is used for outputting the instruction. The instruction may be a control instruction or an adjustment instruction, and the like, without limitation. For example, the second network element obtains a network operation target, and a network index and/or a network state (collectively referred to herein as a network prediction index) of the network operation output by the first network element at a certain time in the future, and combines the network prediction index and the network operation target determining instruction according to the network control model. The network work execution instruction in the network system can change the processing mode of the network to the service state and the data message.

In specific implementation, the second network element may be a network element (such as AMF, SMF, PCF) for processing a service in the network or one of the modules, and may obtain an output of the network data analysis network element, and take a corresponding action to optimize operation of the second network element, thereby optimizing network operation; the second network element may also be an independent network control network element (for example, a network control function (Network Control Function, NCF)), and may analyze the output of the network element according to the network data, decide to take action, control the network elements of one to a plurality of processing services to take corresponding action, and optimize the network operation; the second network element may also be a module in the network data analysis network element itself (such as NWDAF, etc.), or calculate, according to the network data analysis result, the network operation data after the network action is expected to be taken, if the network operation data can reach the target value, the action to be taken is decided, and the network elements that process the service are controlled to take corresponding actions, so as to optimize the network operation. For convenience of description, the following will mainly take an example in which the second network element is a separate network control network element (e.g. NCF).

Alternatively, the first network element and the second network element may be part of a network.

Referring to fig. 5, a flowchart of a model training method according to an embodiment of the present application is taken as an example, where the method is applied to the system shown in fig. 4, and the method includes S101 to S103:

s101, the first network element (such as NWDAF) determines a second network index according to the first network prediction model.

The first prediction model is used for predicting the future network operation state according to the historical and/or current network operation state.

In the embodiment of the application, the network index is used for representing the running state of the network, but the application is not limited to the network index. Correspondingly, the input parameters of the first network prediction model may include a first network index, and the output parameters may include a second network index, where the first network index may include an actual network index of the network at a first time and/or an actual network index before the first time, and the second network index is a network index of the network output by the first network prediction model according to the first network index at a second time, where the second time is later than the first time.

S102, the first network element determines a first prediction deviation.

The first prediction deviation is a deviation of a network index of the network predicted by the first network prediction model at the second moment relative to an actual network index of the network at the second moment. The first predicted deviation is, for example, a deviation of the second network indicator from a third network indicator, wherein the third network indicator is an actual network indicator of the network at the second time instant, such as a network indicator measured at the second time instant.

The network performs network behavior between the first time and the second time, such as the network performs a first instruction to adjust a network parameter of the network. In other words, the third network indicator is the actual network indicator at the second time after the first instruction is executed by the network. It will be appreciated that embodiments of the present application employ execution instructions to represent network behavior, but are not limited thereto, e.g., network behavior may be adjusting network parameters, etc.

In some embodiments, the first prediction bias may be considered a prediction bias of the first network prediction model.

And S103, the first network element adjusts model parameters of the first network prediction model according to the first prediction deviation and the first instruction to obtain a second network prediction model.

Specifically, the first network element adjusts model parameters (such as weight parameters) of the first network prediction model by using a machine learning technology according to the first prediction deviation and the first instruction, so as to obtain an adjusted network prediction model, such as a second network prediction model. The objective of the adjustment is that, when the input parameters of the second network prediction model include the first network index and the first instruction, the deviation (referred to herein as the second prediction deviation) of the network index of the network output by the second network prediction model at the second moment relative to the third network index is smaller than the first prediction deviation.

In some embodiments, the first prediction bias may be considered a prediction bias of the first network prediction model and the second prediction bias may be considered a prediction bias of the second network prediction model. In other words, in the scenario where the network executes the first instruction, the prediction bias of the adjusted network prediction model (i.e., the second network prediction model) is reduced compared to the prediction bias of the network prediction model before adjustment (the first network prediction model).

For example, the first time is time t1, the second time is time t2, and t2=t1+Δt. Taking the network overload level as an example, the actual network overload level of the network at the time t1 is 0 (i.e. the first network index), the network overload level at the time t2 predicted by the first network prediction model is 2 (i.e. the second network index), but the network overload level actually measured at the time t2 is 1 (i.e. the third network index), then the weight parameter of the first network prediction model can be adjusted, so that when the first instruction is executed by the network, the input of the adjusted network prediction model (i.e. the second network prediction model) can embody the influence of the adjustment instruction, and the input parameter of the second network prediction model comprises the actual network overload level 0 of the network at the time t1 and the first instruction, and the prediction result of the second network prediction model can be as close to the network overload level actually measured at the time t2, i.e. the network overload level 1 as possible; when the network does not execute the first instruction, the second network prediction model can still output the prediction result (i.e. the predicted network overload level at the time t2 is 2) according to the same as the original first network prediction model. With such an adjustment, the prediction bias of the second network model in the case of the network executing the first instruction is reduced.

Through the above steps S101 to S103, the training of the network prediction model by using the network behavior (such as the first instruction) as the input parameter is achieved, so that the prediction accuracy of the network prediction model under the network behavior can be rapidly improved, and the network can be rapidly converged to the desired target. Thus, the network can be allowed to realize network intellectualization by using a low-precision network prediction model at first, and the intellectualization of network control can be improved.

In some embodiments, the first network element may cluster the first instructions, so that the prediction bias of the second network prediction model can be reduced compared to the prediction bias of the first network prediction model in a scenario where the network executes the same class of instructions.

It can be understood that the network index change amounts corresponding to different instructions in the same class of instructions (i.e. the change amounts of the network index before and after the execution of the instructions) are all within the same preset range, in other words, the effects of the same class of instructions on the network are similar or similar. Wherein the amount of change may also be referred to as the amount of change. Correspondingly, the first network element adjusts model parameters of the first network prediction model according to the first prediction deviation and the first type of instructions to obtain a second network prediction model.

In specific implementation, the first network element may directly cluster the instructions according to the first instruction, for example, the instructions of a group of similar actions, and may be regarded as a class as long as the difference of specific adjustment parameters is within a certain range; or the first network element clusters the instructions according to the first network index and the second network index, for example, different instructions can be used as the same type as long as the difference of the corresponding network index change amount is within a certain range. If the first instruction belongs to the first type of instruction, the network index change amount corresponding to each instruction in the first type of instruction is within a preset range.

Thus, model training efficiency can be improved.

In some embodiments, the first network element may send the first network index and/or the second network index to the second network element, so that the second network element may provide the first instruction to the first network element according to the first network index and/or the second network index, and the first network element may adjust the first network prediction model according to the first instruction.

The second network element may input the first network index and/or the second network index and the network operation target into the network control model, and the network control model may output a first instruction based on the first network index and/or the second network index and the network operation target, where the first instruction is an instruction that is output by the network control model and is used to make the network index of the network reach the network operation target by using the speculation output by the network control model. It should be understood that the network control model herein may be the control model described above (as in the scheme shown in fig. 3, the network control model used by the central network automation function network element), or may be a network control model described below (as the first network control model or the second network control model in S201 to S203), which is not limited herein.

Therefore, the first network element can obtain the first instruction, and the reliability of the scheme is improved.

In some embodiments, the first network element may perform multiple rounds of training on the network prediction model based on the first instruction (or the first class of instructions), and the adjusted network prediction model obtained after each round of training may be used as an adjusted object (i.e., a network prediction model before adjustment) in the next round of training.

For example, S101 to S103 may be regarded as one training round. In specific implementation, S101 to S103 may be first-round (i.e., first-round) training or non-first-round training, and the following two cases are respectively described:

(1) The first network prediction model is the initial network prediction model when S101 to S103 are the first round. Alternatively, the input of the initial network prediction model may not include the first instruction.

In specific implementation, the first network element (such as NWDAF) can perform offline model training according to the network input/output measurement index acquired in the historical network operation process to obtain an initial network prediction model; in another implementation, the first network element may obtain an initial network prediction model from a model training logic function (Model Training Logic Function, MTLF) network element or the like responsible for model training, and the MTLF network element may perform offline model training using historical data of other similar services and distribute the trained model to the first network element.

Alternatively, after the first round of training, if the second network prediction model does not meet the requirement, for example, the prediction deviation of the second network prediction model exceeds a preset value, a second round of training may be performed, in which the second network prediction model is used as the network prediction model before adjustment. Referring to fig. 6, a flowchart for the second training round includes S104 to S106:

S104, the first network element obtains a fourth network index according to a second network prediction model, wherein the input parameters of the second network prediction model comprise a first network index and a first instruction, and the fourth network index is a network index of the second network prediction model at a second moment (such as t1+delta t) or a third moment (the third moment is later than the second moment, such as t2+delta t, and the application is not limited) according to the first network index and the first instruction;

s105, the first network element determines a second prediction deviation, wherein the second prediction deviation is the deviation of a fourth network index relative to a third network index (namely, when the input parameters of the second network prediction model comprise the first network index and a first instruction, the network index of the network output by the second network prediction model at a second moment is relative to the deviation of the third network index);

S106, the first network element adjusts model parameters of the second network prediction model according to the second prediction deviation and the first instruction to obtain a third network prediction model; when the input parameters of the third network prediction model include the first network index and the first instruction, the deviation (referred to herein as a third prediction deviation) of the network index of the network output by the third network prediction model at the second moment relative to the third network index is smaller than the second prediction deviation. It is understood that the first predictive deviation > the second predictive deviation > the third predictive deviation.

Similarly, if the prediction deviation (e.g., the third prediction deviation) of the third network prediction model meets the requirement, training based on the first instruction is completed, if the prediction deviation of the third network prediction model does not meet the requirement, the process goes to S104, the third network prediction model is used as the network prediction model before adjustment, and the third training is continuously performed with reference to S104-S106, so as to circulate until the network prediction model meeting the requirement is obtained, or until the training termination condition (e.g., the training time reaches the threshold or the training duration exceeds the preset duration, etc.) is met.

(2) If the above-mentioned steps S101 to S103 are not the first round, the first network prediction model is the network prediction model obtained by the previous round of training. In this case, the input of the first network prediction model may further include a first instruction, in other words, the second network index is a network index of the network output by the first network prediction model at the second time when the input parameter of the first network prediction model includes the first network index and the first instruction, and the first control deviation is a deviation of the second network index predicted by the first network prediction model according to the first network index and the first instruction with respect to the third network index. Regarding the case where S101 to S103 are not the first round, reference is made to the description of the flow of fig. 6, which is not expanded here.

Therefore, through multiple rounds of training, the prediction deviation of the network prediction model under the condition that the network executes the first instruction is smaller and smaller, and the precision of the network prediction model is further improved.

In some embodiments, the first network element may train the network prediction model based on other instructions (or other classes of instructions) in addition to the first instructions (or first class of instructions). In one possible design, the first network element triggers training of the network prediction model based on the instructions. For example, the network may periodically generate new instructions at first time intervals, and for each periodically generated instruction, the first network element may train the network prediction model according to the training method described above. Therefore, the prediction precision of the network prediction model obtained by training under various instructions can be improved.

Referring to fig. 7, a flowchart of another model training method according to an embodiment of the present application is taken as an example, where the method is applied to the system shown in fig. 4, and the method includes S201 to S203:

s201, a second network element (such as NCF) determines a first instruction according to a first network control model.

The first network control model is used for presuming network behavior enabling the network to reach a network operation target according to the current network operation state and/or the predicted network operation state. Specifically, the second network element may calculate, according to the first network control model, a network operation state of the network after the network is expected to take a certain network action, and if the network operation state can reach the network operation target, take the network action as a network action for speculatively enabling the network to reach the network operation target.

In the embodiment of the application, the network operation state is represented by a network index, and the network behavior is represented by an instruction, but the application is not limited to the network operation state. Accordingly, the input parameters of the first network control model may include a network operation target and at least one network index, and the output parameters of the first network control model may include a first instruction, where the first instruction is an instruction that is output by the first network control model and is used to make the at least one network index reach the network operation target by using the speculation output by the first network control model, where the first instruction is used to adjust the network parameters of the network.

The network operation target may be implemented in various ways, for example, the specified network performance may be satisfied as the network operation target, and specifically, for example, the network of a certain park is controlled not to be overloaded. Of course, this is merely an example, and is not limited thereto in practice.

S202, the second network element determines a first control deviation.

The first control deviation is a deviation of a behavior effect generated by a first instruction output by the first network control model relative to a behavior effect generated by a network to achieve a network operation target. In the embodiment of the present application, the network indicator change amount may be used to represent the behavior effect, for example, the behavior effect generated by the first instruction is the change amount of at least one network indicator corresponding to the first instruction (i.e., the difference between at least one network indicator before the first instruction is executed by the network and at least one network indicator after the first instruction is executed by the network), the target network indicator change amount capable of achieving the network operation target is the difference between at least one network indicator before the first instruction is executed and the network operation target, and the first control deviation is the deviation of the change amount of at least one network indicator corresponding to the first instruction relative to the target network indicator change amount capable of achieving the network operation target, but is not limited to this in practice.

In some embodiments, the first control deviation may be regarded as a control deviation of the first network control model.

And S203, the second network element adjusts the model parameters of the first network control model according to the first control deviation to obtain a second network control model.

Specifically, the second network element adjusts the first network control model according to the first control deviation and the first instruction by using a machine learning technology, so as to obtain an adjusted network control model, such as a second network control model. Wherein, when the input parameters of the second network control model include a network operation target and at least one network index, a deviation (referred to herein as a second control deviation) of a change amount of the at least one network index corresponding to the second instruction output by the second network control model with respect to a change amount of the target network index is lower than the first control deviation.

In some embodiments, the first control bias may be considered a control bias of the first network control model and the second control bias may be considered a control bias of the second network control model. In other words, in a scenario where it is desired to achieve the network operation target, the control deviation of the adjusted network control model (i.e., the second network control model) is reduced compared to the control deviation of the network control model (the first network control model) before adjustment.

In one implementation, at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of at least one network index is the change amount of a first network index after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the first network indicator (before the network executes the first instruction).

For example, taking the network overload level as an example, the actual network overload level of the current network is 3 (i.e., the first network index), the network overload level that the network expects to reach is 0 (i.e., the network operation target), and then the target network index change amount is-3 (i.e., 3 network overload levels are reduced). The first network control model outputs a first instruction according to a first network index (i.e., network overload level 3) and a network operation target (i.e., network overload level 0), wherein the first instruction is an instruction which is presumed by the first network control model to reduce the network overload level from 3 to 0. After the first instruction is executed by the network, the overload level of the network is measured to be 2, namely, the overload level of the network is reduced by 1 (namely, the change amount of the first network index after the first instruction is executed by the network is-1), the overload level of the network is also required to be reduced by 2 from the network operation target (namely, the first control deviation is (-3) - (-1) = -2), and then the model parameters of the first network control model are adjusted to obtain a second network control model. When the input parameters of the second network control model include the first network index (i.e., network overload level 3) and the network operation target (i.e., network overload level 0), the second network control model outputs a second instruction, and after the second instruction is executed by the network, the network overload level is smaller than 2, for example, 1, that is, the change amount of the first network index after the second instruction is executed by the network is-2 (the network overload level is reduced from 3 to 1), and compared with the change amount (-1) of the first network index corresponding to the first instruction, the change amount (i.e., -2) of the first network index corresponding to the second instruction is closer to the change amount (i.e., -3) of the target network index of the network operation target.

In another implementation manner, at least one network indicator is a second network indicator, where the second network indicator is a network prediction model (the network prediction model may be a network prediction model in the scheme shown in fig. 3, or a first network control model or a second network control model in S101-S103, which is not limited here), and the second time is later than the first time according to an actual network indicator of the network at the first time and/or a network indicator of the network output by an actual network indicator before the first time; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction; the target network indicator change is the difference between the network operating target and the second network indicator (before the network executes the first instruction).

For example, taking the network overload level as an example, the network overload level predicted by the prediction model at the time t2 is 4 (i.e., the second network index), the network expected network overload level is 1 at the time t2 (i.e., the network operation target), and then the target network index change amount is-3 (i.e., the network overload level is reduced by 3). The first network control model outputs a first instruction according to the second network index (i.e. network overload level 4) and the network operation target (i.e. network overload level 1), wherein the first instruction is an instruction which is presumed by the first network control model and enables the network overload level predicted by the prediction model to be 1 at the time t 2. After the first instruction is executed by the network, the prediction model predicts that the network overload level is 3 at the time t2, predicts that the network overload level is reduced by 1 (i.e. the change amount of the second network index is-1 after the first instruction is executed by the network), and further needs to reduce the network overload level by 2 from the target (i.e. the first control deviation is (-3) - (-1) = -2), so that the model parameters of the first network control model are adjusted, and the obtained second network control model is obtained. When the input parameters of the second network control model include a second network index (i.e., network overload level 4) and a network operation target (i.e., network overload level 1), the second network control model outputs a second instruction, and after the second instruction is executed by the network, the prediction model predicts that the network overload level at time t2 is smaller than 3, for example, 2, that is, the change amount of the second network index after the second instruction is executed by the network is-2 (the network overload level is reduced from 4 to 2), and compared with the change amount (-1) of the second network index corresponding to the first instruction, the change amount (i.e., -2) of the second network index corresponding to the second instruction is closer to the target network index change amount (i.e., -3) of the network operation target.

In yet another implementation manner, the at least one network indicator includes a first network indicator and a second network indicator, where the first network indicator is an actual network indicator of the network at a first time and/or an actual network indicator before the first time, and the second network indicator is a network prediction model (the network prediction model may be a network prediction model in the scheme shown in fig. 3, or a first network control model or a second network control model in S101-S103, etc., without limitation), and the second time is later than the first time; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction or the change amount of a first network index after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the first network indicator (before the network executes the first instruction).

It will be appreciated that after the first instruction is executed by the network, it will affect not only the actual network metrics, but also the network metrics predicted by the prediction model (since the prediction model predicts based on the actual network metrics). In some embodiments, the behavior effect of the first instruction on the actual network indicator is the same as or similar to the behavior effect of the first instruction on the predicted network indicator, so that the change amount of at least one network indicator may be the change amount of the second network indicator after the first instruction is executed by the network, or the change amount of the first network indicator after the first instruction is executed by the network.

For example, taking the network overload level as an example, the actual network overload level of the current (time t 1) network is 3 (i.e. the first network index), the network overload level of the predicted time t2 predicted by the prediction model is 4, the network overload level expected to be reached by the network is 0 (i.e. the network operation target), and then the target network index change amount is-3 (i.e. the actual network overload level is reduced by 3 network overload levels). The first network control model outputs a first instruction according to a first network index (i.e., network overload level 3), a second network index (i.e., network overload level 4) and a network operation target (i.e., network overload level 0), wherein the first instruction is an instruction which is presumed by the first network control model to reduce the network overload level from 3 to 0. After the first instruction is executed, the network measures that the overload level of the network is 2, namely the overload level of the network is reduced by 1 (namely the change amount of the first network index after the first instruction is executed by the network is-1), the overload level of the network is required to be reduced by 2 from the network operation target (namely the first control deviation is (-3) - (-1) = -2), the overload level of the network at the time t2 predicted by the prediction model is 3 (namely the change amount of the second network index after the first instruction is executed by the network is-1), and then the model parameters of the first network control model are adjusted to obtain the second network control model. When the input parameters of the second network control model include the first network index (i.e., network overload level 3), the second network index (i.e., network overload level 4), and the network operation target (i.e., network overload level 0), the second network control model outputs a second instruction, and after the second instruction is executed by the network, the network overload level is less than 2, for example, 1, i.e., the change amount of the first network index after the second instruction is executed by the network is-2 (the network overload level is reduced from 3 to 1), and the network overload level at time t2 predicted by the prediction model is 2 (i.e., the change amount of the second network index after the first instruction is executed by the network is-2)). The change amount of the first network index corresponding to the second instruction (i.e., -2) is closer to the target network index change amount of the network operation target (i.e., -3) than the change amount of the first network index corresponding to the first instruction (-1). The change amount of the second network index corresponding to the second instruction (i.e., -2) is closer to the target network index change amount of the network operation target (i.e., -3) than the change amount of the second network index corresponding to the first instruction (-1).

Through the above steps S201 to S203, the second network element may determine the control deviation of the network control model by the amount of change of the network index (such as the amount of change of the first network index or the amount of change of the second network index) before and after the network behavior (such as the first instruction) is executed, and train the network control model based on the control deviation, so that the control accuracy of the network control model may be quickly improved, and the network may be quickly converged to the desired network operation target.

In embodiments of the present application, the second network may initially use a low-accuracy network control model (e.g., the first network control model). The low-precision network control model obtaining mode includes but is not limited to the following modes:

mode 1, the second network element (e.g., NCF) may construct initial rules for the network control model based on standard defined behavior logic (e.g., PCF reduces aggregate maximum bit rate authorized for the terminal (AAGGREGATE MAXIMUM BIT RATE, AMBR) may reduce network overload, etc.);

Mode 2, a second network element (e.g., NCF) may develop initial rules for constructing a network control model based on expert experience in historically operating and maintaining campus network formation.

In some embodiments, each rule includes a network predictor and a corresponding adjustment action (or adjustment instruction), wherein the magnitude of the adjustment action employs an empirically initialized parameter. For example, the following table 1 is generated according to expert experience:

Table 1 initial rules of network control model

It will be appreciated that table 1 is exemplified by network operation targets to prevent network overload (i.e., network overload level 0), and that the second network element is exemplified by NCF, but is not limited thereto.

It should be understood that table 1 is only one example, and is not limited thereto in practice.

In a possible implementation manner, the second network element may further send a first instruction to the first network element, where the first network element operates with the network prediction model, and after the first network element receives the first instruction, model parameters of the network prediction model that operates on the first network element may be adjusted based on the first instruction (see, in particular, the descriptions of S101 to S103 above).

In some embodiments, the second network element may perform multiple rounds of training on the network control model based on the first instruction, and the network control model after each round of training may be used as the adjusted object (i.e., the network control model before adjustment) in the next round of training.

For example, the above-described S201 to S203 may be regarded as one round of training. If the second network control model does not meet the requirement, for example, the control deviation of the second network control model exceeds a preset value, performing the next training, taking the second network control model as the network control model before adjustment in the next training, and referring to S201-S203, the flow of the next training can be referred to, and no more description is needed here, so as to circulate until the network control model meeting the requirement is obtained, or until the training termination condition (for example, the training frequency reaches a threshold value or the training duration exceeds the preset duration, etc.) is met.

Therefore, through multiple rounds of training, the control deviation of the network control model in the scene of executing the first instruction by the network can be smaller and smaller, and the accuracy of the network control model is improved.

In some embodiments, the second network element may train the network control model based on other instructions in addition to the first instructions.

In one possible design, the first network element may trigger training of the network control model based on the instructions. For example, new instructions are periodically generated in the network at second time intervals, for each of which the second network element can train the network control model according to the training method described above.

Wherein the first time interval and the second time interval are the same or different, in other words, the instruction used by the second network element to train the network control model and the instruction used by the first network element to train the network prediction model may be the same or different.

Therefore, the control precision of the network control model obtained by training under various instructions can be improved.

The above description is provided for the training method of the network prediction model and the training method of the network control model. It should be noted that, the training method of the network prediction model and the training method of the network control model may be implemented separately or in combination with each other, and the present application is not limited thereto.

For example, fig. 8A and fig. 8B are exemplary diagrams of another model training method according to an embodiment of the present application, where fig. 8A is a logic block diagram of a network control system performing model training, and fig. 8B is a flow chart of the network control system performing model training. In this example, the first network element is exemplified by NWDAF and the second network element is exemplified by NCF. The specific flow comprises the following steps:

initial state:

S300a, NWDAF obtains an initial network prediction model.

The method for obtaining the initial network prediction model may be referred to above, and will not be described herein.

S300b, NCF obtains initial network control model.

The method for obtaining the initial network control model may refer to the above, and will not be described herein.

S301, the NCF acquires a network operation target. For example, depending on business experience needs, it may be desirable to control the overload level of the campus network to less than 1.

After the NCF acquires the network operation target, the system enters an operation state.

Operating state:

In an operating state, the system may periodically perform the following steps with a fixed time interval Δt (e.g., 1 minute):

optionally, during system operation, the NCF may also acquire updated network operation targets (e.g., added, deleted, modified network operation targets), at which point the operating state may be re-entered.

S302a, NWDAF may acquire the network measurement index of the current time (i.e., the actual network index of the current time) from AMF, SMF, UPF, PCF or the like as input information for performing network prediction.

For example, NWDAF obtains the average packet delay (AVERAGE PACKET DELAY), average packet loss rate (Average Loss Rate), and Throughput (Throughput) of each terminal in the campus at the historical and current time (e.g., time t) for determining the data network performance analysis (DN Performance Analytics).

S302b, optionally, the NCF may obtain a network measurement indicator at the current time (i.e. an actual network indicator at the current time) from NWDAF, and determine the current network overload level (e.g. overload level 0); or the NCF may obtain data network performance analysis results from NWDAF to determine that the current network overload level is 0.

S303, NWDAF determines, according to the network measurement indexes at the current time (e.g., time t) and the history, the network prediction index (i.e., the network index predicted by the initial network prediction model) using the network prediction model obtained in S300a, for example, the network prediction index is a network overload level of 2, where the future Δt duration may occur.

S304, the NCF obtains a network prediction index from NWDAF (e.g., the future Δt duration network overload level is predicted to be 2).

S305, the NCF determines a network adjustment instruction (e.g., a first instruction) according to the initial network control model obtained in S300 b.

There are various implementations of the NCF determining the network adjustment instructions from the initial network control model, and the following are examples of several possible implementations:

In mode 1, the NCF may determine a network adjustment instruction (e.g., a first instruction) according to the network measurement index at the current time (e.g., time t) obtained in step S302b, in combination with the initial network control model obtained in step S300b, and after the NCF speculates that the network adjustment instruction is executed, the network measurement index may reach the network operation target obtained in step S301.

In mode 2, the NCF may determine a network adjustment instruction (e.g., a first instruction) according to the network prediction index obtained in S303 (e.g., the future Δt duration network overload level is predicted to be 2), and in combination with the initial network control model obtained in S300b, the NCF speculatively executes the network adjustment instruction to enable the network prediction index to reach the network operation target obtained in step S301.

In mode 3, the NCF may determine a network adjustment instruction (e.g., a first instruction) according to the network prediction index (e.g., the network overload level of the future Δt duration is predicted to be 2) obtained in S303, and the current network measurement index (e.g., the overload level of the time t is 0) obtained in S302b, and in combination with the initial network control model obtained in S300b, the NCF speculatively executes the network adjustment instruction, so that the network measurement index may reach the network operation target obtained in step S301.

For example, the NCF determines to obtain a user plane data congestion "User Data Congestion" analysis from NWDAF and reduces the authorized Session MBR for the PDU Session for the one application with the largest traffic in the "User Data Congestion" forecast to 75% of the original. The NCF then obtains a "User Data Congestion" analysis from NWDAF in a manner similar to S302b, listing several applications (not shown) with maximum traffic in the campus network. The NCF then determines to reduce the authorized Session MBR for all PDU sessions of one application where traffic is greatest to 75% of the original, and the NCF also determines to reduce the authorized AMBR for low priority terminals on the campus to 60% of the original, generating corresponding network adjustment instructions.

S306a, NCF sends a network adjustment instruction to the service processing network element (e.g. AMF, SMF, UPF, PCF, etc.).

And S306b, the NCF sends a network adjustment instruction to the NWDAF network element, and the network adjustment instruction is used for informing NWDAF that the network adjustment instruction is executed to influence the network operation when the subsequent predictive analysis is output.

S307, at least one network function adjusts network operation according to the network adjustment instruction.

For example, the PCF modifies the Session management Policy (Session Management Policy, SM Policy) specifying all PDU sessions to which the application corresponds, reduces the authorization Session MBR to 75% of the original, and sends the updated SM Policy to the SMF. After the SMF receives the message, the UPF is notified to limit all PDU sessions corresponding to the appointed application to 75% of the original maximum data rate. The PCF modifies the access management Policy (ACCESS MANAGEMENT Policy, AM Policy) to reduce the authorized AMBR of the low priority terminal in the campus to 60% of the original, and sends the modified authorized AMBR of the low priority terminal to the gNB through the AMF, and the gNB limits the aggregation rate of the terminals according to the rate of 60% of the original AMBR for the low priority users.

S308a, after a time interval Deltat (i.e. at the time t+Deltat), NWDAF obtains the current network measurement index (i.e. after the execution of the network adjustment instruction), and the specific method refers to S302a.

S308b, NCF may obtain a network measurement indicator at the current time (i.e., time t+Δt) from NWDAF, or obtain a network prediction indicator at the next time (e.g., time t+2Δt) from NWDAF.

For example, when the NCF determines the network adjustment instruction in the mode 1 or the mode 3 in S305, the NCF may acquire the network measurement indicator (such as overload level 1) at the current time (i.e. time t+Δt) from NWDAF; for example, when the NCF determines the network adjustment command in the mode 2 in S305, the NCF may obtain the network prediction index (e.g. overload level 1) at the current time (i.e. time t+Δt) from NWDAF.

It will be understood that the network prediction index at the current time (i.e. t+Δt time) refers to the network index of the network predicted at the current time at the next time (i.e. t+2Δt time).

S309a, NWDAF determines a prediction deviation of the network prediction model under the network adjustment instruction according to the network adjustment instruction obtained in S306b, the network measurement index obtained in S308a, and the network prediction index obtained in S303.

For example, if the overload level predicted in S304 is 2, but the overload level actually measured at the corresponding time of prediction in S308a is 1, the prediction deviation is 1-2= -1.

S309b, NCF determines the behavioral effect of the network adjustment instruction of S306 b.

It will be appreciated that the behavior effect of the network adjustment instruction in S309b may be determined differently according to the different determination manners of the network adjustment instruction in S305.

When the network adjustment instruction is determined in the above-described mode 1 or mode 3, the NCF may determine the behavioral effect of the network adjustment instruction of S306b based on the network measurement index obtained in S308b and the network measurement index obtained in S302 b.

For example, the change amount (i.e., the change amount of the network measurement index before and after the execution of the network adjustment instruction) of the network measurement index at the current time (i.e., the time t+Δt) relative to the network index at the previous time (i.e., the time t) is determined (i.e., the overload level is reduced by 1 level), specifically, for example, the network overload level of the park at the current time (i.e., the time t+Δt) is 1, the average packet delay (AVERAGE PACKET DELAY) of each terminal, the average packet loss rate (Average Loss Rate) and the change condition of the Throughput (Throughput) at the current time (i.e., the time t+Δt) relative to the previous time (i.e., the time t).

When the network adjustment instruction is determined in the above manner 2, the NCF may determine the behavioral effect of the network adjustment instruction of S306b according to the comparison of the network prediction index obtained in S308b and the network prediction index obtained in S304.

For example, the change amount of the network prediction index at the current time (i.e. the time t+Δt) relative to the network prediction index at the previous time (i.e. the predicted overload level is reduced by 1 level), that is, the change amount of the network prediction index before and after the execution of the network adjustment instruction is determined, specifically, for example, the average packet delay (AVERAGE PACKET DELAY) of each terminal predicted at the current time (i.e. the time t+Δt), the change condition of the average packet loss rate (Average Loss Rate) and the Throughput (Throughput) at the current time (i.e. the time t+Δt) relative to each corresponding index predicted at the previous time (i.e. the time t).

The NWDAF performs online learning adjustment on the network prediction model according to the prediction deviation determined in S309a and the network adjustment instruction in S306b, and adjusts the network prediction model such that the prediction deviation under a certain type of adjustment instruction (i.e., the category corresponding to the network adjustment instruction in S306 b) is reduced, that is, adjusts the parameters of the network prediction model in the direction in which the prediction deviation under a certain type of adjustment instruction is reduced.

For example, under the network measurement index and the operation trend at time t, the predicted network overload level may be lowered by one level like the network adjustment instruction of S306 b. Then, adjusting the weight parameters of the network prediction model, enabling the input of the network prediction model to embody the influence of an adjustment instruction, and correcting the prediction to reduce the result by one level under the network adjustment instruction similar to the S306 b; when no network adjustment instruction is executed, the network adjustment instruction is output according to the original prediction result. Through such online learning, in the case where the network is caused to execute the network adjustment instruction like S306b, the prediction bias of the network prediction model is reduced.

And S310b, NCF performs online learning according to the behavior effect (such as the change amount of the network measurement index before and after the execution of the network adjustment instruction) determined in S309b and the deviation of the behavior effect generated by the network adjustment instruction relative to the behavior effect generated by the network to reach the network operation target (such as the deviation of the actual network index relative to the network operation target at the corresponding time (i.e. at the time t+Δt), specifically, for example, the actual overload level is still 1 level higher than the operation target after the execution of the network adjustment instruction), and adjusts the weight parameters of the network control model so as to reduce the control deviation.

For example, the NCF would adjust the reduction of the authorized Session MBR for the PDU Session of the one application with the largest traffic in the current "User Data Congestion" analysis to the original 75% to the original 60% and the reduction of the authorized AMBR for the low priority terminal in the campus to the original 60% to the original 40% when the overload level is raised by 2 levels for the predicted Δt duration.

It can be understood that, after updating the network prediction model and the network control model, NWDAF and the NCF can immediately execute S303 to S307 at the current time (t+Δt), and execute S308a to S310b as described above when the next time (t+2Δt) arrives, and thus the network updates the network adjustment command once every Δt time period; NWDAF and NCF may not execute S303 to S307 at the current time (t+Δt), but wait until the next time (t+2Δt), and NWDAF and NCF execute S302a to S310b described above using the updated network prediction model and the network control model, respectively, so that the network may update the network adjustment instruction every 2 Δt time period. The period of model training is not particularly limited in the embodiment of the present application.

In the scheme, the determined network adjustment instruction is used as the input of the network prediction model, so that the network prediction model can fully embody the influence of the network adjustment instruction on the network operation, and further the prediction deviation is reduced. In addition, the network control model is adjusted according to the change of the network measurement index before and after the execution of the network adjustment instruction and the deviation of the actual network index relative to the operation target at the corresponding moment, so that the control deviation is reduced and the operation target is converged more quickly.

In some embodiments, the initial network control model may be determined based on network intent. For example, the initial value of the model parameter of the first network control model is determined according to a network intention, where the network intention may be an intention (or a requirement or an expectation or a target, etc.) of the network operator input by the network operator on the network, and the network element in the network control system may translate the network intention into a network operation target and a set of network adjustment rules, where each rule in the set of network adjustment rules is an adjustment action corresponding to the network prediction index, and an amplitude of the adjustment action adopts an empirical parameter. The network element responsible for the intent translation may be a module of a network element or a network control network element (Network Control Function) that processes traffic in the network, or may be an independent intent driver function (Intention Driven Function, IDF) network element (i.e., dedicated to translating the network intent into a network operation target and a network adjustment rule), etc., which is not limited in this regard by the present application.

For a better understanding of the above, a detailed example is presented herein.

For example, fig. 9A and fig. 9B are exemplary diagrams of another model training method according to an embodiment of the present application, where fig. 9A is a logic block diagram of a network control system performing model training, and fig. 9B is a flow chart of the network control system performing model training. In this example, the first network element is NWDAF, the second network element is an NCF, and an independent IDF is further included to translate network intent and pass the translation result to the NCF. The specific flow comprises the following steps:

initial state:

S400a, NWDAF obtains an initial network prediction model.

The method for obtaining the initial network prediction model may be referred to above, and will not be described herein.

S400b, IDF translates the intent entered by the network operator into a network operation target and a set of network adjustment rules. The network adjustment rule comprises a network prediction index combination and a corresponding adjustment action, wherein the adjustment amplitude can adopt parameters in the adjustment rule as initialization parameters.

For example, an intent to describe as "preventing a network from entering an overload state" may be translated into:

(1) Network operation target: overload class <1;

(2) Network adjustment rules: the traffic of the campus network is reduced according to the predicted overload levels as per table 2 below.

Table 2 network adjustment rules

It should be understood that table 2 is only one example, and is not limited thereto in practice.

S400c, NCF obtains the network adjustment rule from the IDF.

And S400d, initializing a network control model by using the NCF by using a network adjustment rule to obtain an initial network control model.

S401, the NCF acquires a network operation target from the IDF.

For example, depending on business experience needs, it may be desirable to control the overload level of the campus network to less than 1. After the NCF obtains the operation target, the system enters an operation state.

Operating state:

In an operating state, the system may perform the following steps with a fixed time interval Δt (e.g., 1 minute):

Alternatively, during system operation, the NCF may update the network control model if updated network adjustment rules are obtained from the IDF. The operational state may be re-entered if the NCF obtains updated network operational objectives from the IDF.

S402, NWDAF may obtain, from AMF, SMF, UPF, PCF or the like, network measurement indexes at the current time (e.g., time t) (i.e., network measurement indexes before instruction execution) as input information for performing network prediction. Specific implementation of this step may refer to S302a.

S403, NWDAF, according to the obtained network measurement index of the history and the current time (such as time t), determine a network prediction index (i.e. a network index predicted by the initial network prediction model) by using the network prediction model obtained in S400 a. For example, the network overload level that may occur for the future Δt duration is 2 for the network predictor.

S404, the NCF obtains a network prediction index from NWDAF (e.g., the future Δt duration network overload level is predicted to be 2).

S405, the NCF determines a network adjustment instruction (such as the first instruction above) according to the network prediction index (for example, the future Δt duration network overload level is predicted to be 2) in combination with the initial network control model obtained in S400 d.

For example, the NCF determines to reduce the grant AMBR of the low priority terminal to 60% of the original, and to reduce the grant Session MBR of the PDU Session of the non-critical provisioning application to 75% of the original.

S406a, NCF sends a network adjustment instruction to the service processing network element (e.g., AMF, SMF, UPF, PCF, etc.).

And S406b, the NCF sends a network adjustment instruction to the NWDAF network element, and the network adjustment instruction is used for informing NWDAF that the network adjustment instruction is executed to influence the network operation when the subsequent predictive analysis is output.

S407, at least one network function adjusts network operation according to the network adjustment instruction. Reference is made in particular to S307 above.

S408a, after a time interval Δt (i.e. at time t+Δt), NWDAF obtains the current network measurement index (i.e. after the network adjustment instruction is executed), and the specific implementation method may refer to S402.

S408b, NCF may obtain a network measurement indicator (e.g. overload level 1) at the current time (i.e. time t+Δt, i.e. the time after the execution of the network adjustment instruction) from NWDAF.

And S409a and NWDAF determine the prediction deviation of the network prediction model under the network adjustment instruction according to the network adjustment instruction obtained in S406b, the network prediction index obtained in S403 and the network measurement index obtained in S408a (i.e. the network measurement index after the instruction is executed).

For example, the overload level predicted in S404 is 2, but the overload level actually measured in S408a at the predicted corresponding time is 1.

S409b, NCF determines, according to the network measurement index obtained in S408b, the behavior effect of the network adjustment instruction in S406b (e.g., determines the change amount (i.e., the overload level is reduced by 1 level) of the network measurement index at the current time (i.e., time t+Δt) relative to the network index at the previous time (i.e., time t)), i.e., the change amount of the network measurement index before and after the execution of the network adjustment instruction.

S410a and NWDAF perform online learning adjustment on the network prediction model according to the prediction bias determined in S409a and the network adjustment instruction of S406 b. Specific methods may refer to S310a.

The step 410b, the NCF performs online learning according to the behavior effect determined in the step 409b (i.e., the amount of change of the network measurement index before and after the execution of the network adjustment instruction, such as that the overload level is reduced by 1), and the deviation of the network index predicted by the network prediction model relative to the network operation target (i.e., the deviation of the behavior effect generated by the network adjustment instruction relative to the behavior effect generated by the network to reach the network operation target, such as that the overload level needs to be reduced by 2), so as to adjust the weight parameters of the control model, so that the deviation of the network index predicted by the network prediction model relative to the network operation target is reduced, that is, the behavior effect of the network adjustment instruction can be more matched with the deviation of the network prediction index relative to the network operation target.

For example, the NCF may adjust the reduction of the authorization Session MBR for a PDU Session of a non-critical care application to 75% to 60% and the reduction of the authorization AMBR for a low priority terminal in the campus to 60% to 40% when the overload level reaches 2 for the predicted Δt time period.

At the next time (t+2Δt), NWDAF and NCF may perform the above-described S402 to S410b using the updated network prediction model and network control model, respectively, and so on.

In the scheme, the system can initialize the network control model according to the coarse-precision intention translation result, so that the difficulty in realizing the intention translation is reduced. The network prediction model and the network control model are subjected to linkage adjustment, so that the behavior effect of the network adjustment instruction is closer to the deviation of the network prediction index relative to the network operation target, the behavior effect is more matched with the deviation of the predicted network index relative to the network operation target, and the cooperation of prediction and control can be more accurate.

In some embodiments, in order to achieve the objective of service guarantee for a part of users, the network control system may further include a third network element and a fourth network element, where the third network element is configured to output a network prediction index (abbreviated as a user prediction index) corresponding to a target user, the fourth network element is configured to output a network adjustment instruction (abbreviated as a user adjustment instruction) corresponding to the target user, and the first network element is configured to output a network prediction index of the whole network, and the second network element is configured to output a network adjustment instruction (including a user adjustment instruction) of the whole network.

The first network element may send a network prediction index (e.g., a second network index) of the network entity to the third network element, where the third network element may run the user prediction model, and the third network element may output the user prediction index (e.g., a network index corresponding to a target user predicted by the user prediction model at the second time) by using the network prediction index of the network entity as an input of the user prediction model, and may provide the user prediction index to the fourth network element. And the fourth network element can operate the user control model, takes a user prediction index, a network operation target (simply called a user operation target) corresponding to a target user and the like as input of the user control model, outputs a network adjustment instruction (namely a user adjustment instruction) corresponding to the target user, and transmits the user adjustment instruction to the second network element. When the second network element outputs the network adjustment instruction based on the network control model (such as the first network control model), the second network element can also take the user adjustment instruction as an input parameter, so that the network adjustment instruction (such as the first instruction) output by the network control model comprises the user adjustment instruction, and further, the service experience of the target user is ensured.

It can be appreciated that the third network element may be another network entity different from the first network element (e.g., the third network element and the first network element are two independent NWDAF respectively), and the third network element may be integrated with the first network element in the same network entity (e.g., the third network element and the first network element are different functional modules in the same NWDAF), which is not a limitation of the present application. The fourth network element may be another network entity different from the second network element (e.g., the fourth network element and the second network element are PCF and NCF, respectively), or the fourth network element may be integrated with the first network element in the same network entity (e.g., the fourth network element and the second network element are different functional modules in the same PCF).

Further, the network control system may include an IDF (abbreviated as a network IDF) corresponding to a network (for translating a network intention into a network operation target and a network adjustment rule), and may further include an IDF (abbreviated as a user IDF) corresponding to a user (for translating a user intention into a user operation target and a user adjustment rule). Similarly, the network IDF and the user IDF may be two different network entities, or may be different functional modules in the same network entity, which is not limited by the present application.

For a better understanding of the above, a detailed example is presented herein.

For example, fig. 10A and fig. 10B are schematic diagrams illustrating another model training method according to an embodiment of the present application, where fig. 10A is a logic block diagram of a network control system performing model training, and fig. 10B is a flow chart of a network control system performing model training. The specific flow comprises the following steps:

initial state:

S500a, network NWDAF, and user NWDAF obtain initial predictive models (i.e., an initial network predictive model and an initial user predictive model), respectively.

S500b to S500d, reference may be made to S400b to S400d.

S500e, the user IDF translates the user intention into a user operation target and a set of user adjustment rules, wherein each user adjustment rule comprises an adjustment action corresponding to a user prediction index, and the amplitude of the adjustment action can adopt experience parameters.

For example, the user intent is described as "guarantee the business experience of the target user", which the user IDF can translate into:

(1) The target user network operates the target: the traffic experience MOS > =4.0 for the target user;

(2) User adjustment rules: and according to the following table 3, according to the user prediction service experience index, the bit rate and the QoS priority level of the target user are improved.

TABLE 3 user adjustment rules

It should be understood that table 3 is only one example, and is not limited thereto in practice.

S500f, PCF obtains user adjustment rules from the user IDF.

S500g, the PCF uses the user adjustment rule to initialize the user control model corresponding to the target user, and an initial user control model is obtained.

S501a, NCF obtains a network operation objective from the network IDF, such as controlling the overload level of the campus network to be less than 1.

S501b, the PCF obtains a user operation target from the user IDF, for example, the service experience MOS > =4.0 of the target user.

After the network operation target and the user operation target are obtained, the system enters an operation state.

Operating state:

In an operating state, the system may perform the following steps with a fixed time interval Δt (e.g., 30 seconds):

alternatively, during system operation, the operating state may be re-entered if the NCF obtains updated network adjustment rules or updated network operating targets from the network IDF, or the PCF obtains updated user adjustment rules or updated user operating targets from the user IDF.

S502a, the network NWDAF may obtain network measurement indexes of the history and the current time (e.g., time t) from AMF, SMF, UPF, PCF, etc., as input information for performing the network prediction model.

S502b, the user NWDAF may acquire the measurement index of the target user at the current time (e.g., time t) from AMF, SMF, UPF, PCF, or may acquire the measurement index of the target user from the network NWDAF, as input information of the user prediction model.

S503, the network NWDAF determines, according to the network measurement indexes of the acquired history and the current time (such as time t), the network prediction index using the network prediction model acquired in S500a, for example, the network overload level that may occur in the future Δt minutes is 2.

S504, NCF obtains network prediction index from network NWDAF; the user NWDAF obtains the network predictors from the network NWDAF as input to the focused user predictive analysis.

S505, the user NWDAF determines a user prediction index of the target user (for example, the service experience MOS value of the target user in the future delta t minutes is 3.3) according to the user measurement index of the acquired history and the current moment (for example, the moment t) and the network prediction index of the network NWDAF by using the user prediction model acquired in S500 a.

S506, the PCF obtains a user prediction index of the target user from the user NWDAF (for example, the service experience MOS value of the target user is 3.3 in the future delta t minutes).

S507, the PCF determines a user adjustment instruction (or a user adjustment parameter) of the target user according to the user prediction index obtained in S506 by using the user control model obtained in S500 g. For example, the future Δt minutes target user's service experience MOS value 3.3 matches the MOS <3.5 condition, the maximum bit rate of the target user's PDU session needs to be set to 50Mbps, and two QoS parameters, namely, the delay and the packet loss rate of the target user are reduced, which are 30ms and 1/10000, respectively;

And S508, the PCF sends a user adjustment instruction of the target user to the user NWDAF and simultaneously sends the user adjustment instruction to the NCF.

And S509, the NCF determines a network adjustment instruction according to a network prediction index (for example, the future delta t minute network overload level is predicted to be 2) and the network control model obtained in the S500 d. For example, the NCF determines to reduce the grant AMBR of the low priority terminal to 60% of the original, and to reduce the grant Session MBR of the PDU Session of the non-critical provisioning application to 75% of the original.

It will be appreciated that the NCF may accommodate the user adjustment instruction determined by the PCF in determining the network adjustment instruction, or may modify a portion of the user adjustment instruction that conflicts with a portion of the network adjustment instruction of the entire network. The NCF may take the received or modified user adjustment instructions as part of the finalized network adjustment instructions.

S510a, NCF sends the network adjustment instruction determined in S510 to the service processing network element (e.g., PCF, SMF, AMF, UPF, etc.), where the network adjustment instruction includes a user adjustment instruction.

S510b, the NCF sends the network adjustment instruction determined in S510 to the network NWDAF element, to notify the network NWDAF that the subsequent prediction analysis output considers the effect of the executing network adjustment instruction on the network operation.

And S511, each network function adjusts the network operation according to the network adjustment instruction obtained in the S510 a.

S512a, after a time interval Δt, i.e. at time t+Δt, the network NWDAF obtains the current (i.e. after the instruction is executed) network measurement index. Such as bit rate, packet Wen Shiyan, packet loss rate, overload level, etc.

S512b, NCF may obtain the network measurement indicator at the current time (i.e. t+Δt) from the network NWDAF, and may also determine the change amount of the network indicator at the current time (i.e. t+Δt) relative to the previous time (i.e. t) (e.g. the actually measured overload level is reduced by one level).

S512c, the user NWDAF may obtain the network measurement index at the current time (i.e., t+Δt) from the network NWDAF.

S513, the user NWDAF calculates the MOS value (for example, 3.9) of the target user according to the network measurement index after the adjustment instruction is executed, and further calculates the user measurement index change amount corresponding to the target user after the instruction is executed, for example, the MOS value of the target user is increased by 0.6 from the original predicted value. User NWDAF sends the user measurement index change to the PCF.

S514a, the PCF determines the behavior effect of the user adjustment instruction according to the network measurement index change amount obtained in the S513 (for example, the MOS value of the user is improved by 0.6 from the original predicted value); the PCF performs online learning according to the behavior effect of the target user and the deviation of the user prediction index predicted by the user prediction model relative to the user operation target (for example, the MOS value of the target user service experience is still 0.1 worse than that of the target 4.0), and adjusts the weight parameters of the user control model, so that the deviation of the user prediction index predicted by the user prediction model relative to the user operation target is reduced, that is, the behavior effect of the target user is more matched with the deviation of the user prediction index predicted by the user prediction model relative to the user operation target (for example, the adjusted delay QoS parameter is modified from 30ms to 20 ms).

S514b, a user NWDAF determines a prediction deviation under the user adjustment instruction (namely, the user MOS value is improved by 0.6 from the original prediction value) according to the user adjustment instruction obtained in S508 and the user measurement index obtained in S512 c; and the user NWDAF performs online learning adjustment on the user prediction model according to the prediction deviation and the user adjustment instruction, so that the prediction deviation of the user prediction model is reduced. Specific adjustment methods are given for adjustment methods that can refer to the above network prediction model.

And S515a, the network NWDAF performs online learning adjustment on the network prediction model according to the prediction deviation determined in the S512b and the network adjustment instruction of the S510 a. Specific adjustment methods are given for adjustment methods that can refer to the above network prediction model.

And S515b, NCF determines the behavior effect of the network adjustment instruction according to the change amount of the network measurement index obtained in S512b, namely, the change condition of the network overload level of the park (namely, the overload level is only reduced by 1), and the deviation of the network prediction index predicted by the network prediction model relative to the network operation target (namely, the deviation of the behavior effect generated by the network adjustment instruction relative to the behavior effect generated by the network to reach the network operation target, namely, the overload level needs to be reduced by 2), performs online learning on the network control model, adjusts the weight parameter of the network control model, and enables the deviation of the network prediction index predicted by the network prediction model relative to the network operation target to be reduced, namely, enables the behavior effect of the network instruction output by the network control model to be more matched with the deviation of the network prediction index relative to the network operation target. For example, the NCF would adjust the grant Session MBR for a PDU Session for a non-critical care application to decrease to 75% of the original to 60% when the overload level reaches 2 within a predicted Δt minutes. Reducing the authorized AMBR of the low priority terminals on the campus to 60% of the original is adjusted to 40% of the original.

At the next time (t+2Δt), the network NWDAF, NCF, subscriber NWDAF, PCF, and the like may perform the above-described S502a to S515b using the updated model, and so on.

In the above scheme, the output parameter (i.e. the network prediction index) of the network prediction model can be used as the input parameter of the user prediction model, so that the user prediction model can fully consider the potential change of the network when outputting the user prediction network index, and the accuracy of the user prediction model can be improved. The network adjustment instruction (namely, the user adjustment instruction) corresponding to the target user is used as the input of the network control model, so that the network control model can coordinate the decision conflict between the target user and the network, fully consider the influence of the adjustment of the target user on the network operation, further improve the accuracy of the network control model, and consider the overall operation performance of the network and the service experience of the target user.

It will be appreciated that the implementation manners of the above embodiments may be implemented separately or in combination with each other, and the present application is not limited thereto.

Fig. 11 and 12 are schematic structural diagrams of a possible communication device according to an embodiment of the present application. These communication devices may be used to implement the functions of the first network element or the second network element in the above method embodiments, so that the beneficial effects of the above method embodiments may also be implemented. In the embodiment of the present application, the communication device may be a first network element or a second network element, or may be a module, a unit, or a technical means (such as a chip) applied to the first network element or the second network element.

As shown in fig. 11, the communication device 1100 includes a determination module 1110 and a training module 1120. The communication device 1100 is configured to implement the functions of the first network element or the second network element in the above-described method embodiment.

In a first embodiment, the communication device is configured to implement the function of the first network element in the foregoing method embodiment, and the determining module 1110 is configured to determine, according to a first network prediction model, a second network indicator, where an input parameter of the first network prediction model includes the first network indicator, the first network indicator includes an actual network indicator of the network at a first time and/or an actual network indicator before the first time, and the second network indicator is a network indicator of the network output by the first network prediction model according to the first network indicator at a second time, where the second time is later than the first time;

The training module 1120 is configured to determine a first prediction bias, where the first prediction bias is a bias of the second network indicator relative to a third network indicator, the third network indicator is an actual network indicator of the network at a second time, and the network executes a first instruction between the first time and the second time, where the first instruction is used to adjust a network parameter of the network; according to the first prediction deviation and the first instruction, model parameters of the first network prediction model are adjusted, and a second network prediction model is obtained; when the input parameters of the second network prediction model comprise a first network index and a first instruction, the deviation of the network index of the network output by the second network prediction model at the second moment relative to the third network index is smaller than the first prediction deviation.

In one possible design, training module 1120 is to: clustering the instructions according to the first instructions or clustering the instructions according to the first network index and the second network index to obtain first-class instructions; wherein the first instruction belongs to a first class of instructions; and adjusting model parameters of the first network prediction model according to the first prediction deviation and the first type instruction to obtain a second network prediction model.

In one possible design, the network indicator change amount corresponding to each instruction in the first type of instruction is within a preset range.

In one possible design, the communication device may further include: a transceiver module 1130, configured to send the first network indicator and/or the second network indicator to the second network element; the method comprises the steps of receiving a first instruction from a second network element, wherein the first instruction is output by a network control model running on the second network element, and input parameters of the network control model comprise a first network index and/or a second network index and further comprise a network running target.

In one possible design, the network operation target may include a network operation target corresponding to the target user, and the first instruction may include a network adjustment parameter corresponding to the target user.

In one possible design, transceiver module 1130 is also configured to: and sending a second network index to the third network element, wherein the second network index is used for outputting the network index corresponding to the target user at the second moment by the user prediction model.

In one possible design, the input parameters of the first network prediction model may further include a first instruction, and the second network index is a network index of the network output by the first network prediction model at the second moment according to the first network index and the first instruction.

In a second embodiment, the communication device is configured to implement the function of the second network element in the foregoing method embodiment, and the determining module 1110 is configured to determine the first instruction according to the first network control model; the input parameters of the first network control model comprise a network operation target and at least one network index, the first instruction is an instruction which is output by the first network control model and is used for speculatively enabling the at least one network index to reach the network operation target, and the first instruction is used for adjusting the network parameters of a network; the training module 1120 is configured to determine a first control deviation, where the first control deviation is a deviation of an amount of change of at least one network indicator corresponding to the first instruction relative to an amount of change of a target network indicator that can reach a network operation target; adjusting model parameters of the first network control model according to the first control deviation to obtain a second network control model; when the input parameters of the second network control model comprise a network operation target and at least one network index, the deviation of the change amount of at least one network index corresponding to the second instruction output by the second network control model relative to the change amount of the target network index is lower than the first control deviation.

Illustratively, the at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of at least one network index is the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index.

The at least one network index is a second network index, wherein the second network index is a network index of the network at a second moment, and the second moment is later than the first moment, according to an actual network index of the network at the first moment and/or a network index of the network output by an actual network index before the first moment; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the second network index.

The at least one network index comprises a first network index and a second network index, wherein the first network index is an actual network index of the network at a first moment and/or an actual network index before the first moment, and the second network index is a network index of the network, output by the network prediction model according to the first network index, at a second moment, and the second moment is later than the first moment; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction or the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index.

In one possible design, initial values of model parameters of the first network control model may be determined based on network intent.

In one possible design, the communication device may further include: the transceiver module 1130 is configured to send a first instruction to a first network element, where the first instruction is used to adjust a model parameter of a network prediction model running on the first network element.

In one possible design, the network operation target may include a network operation target corresponding to the target user, the first instruction may include a network adjustment parameter corresponding to the target user, and the input parameter of the first network control model may further include a network adjustment parameter corresponding to the target user; the communication device may further include a transceiver module 1130 configured to obtain, from the fourth network element, a network adjustment parameter corresponding to the target user, where the network adjustment parameter is output by a user control model running on the fourth network element.

The more detailed descriptions of the determining module 1110, the training module 1120, and the transceiver module 1130 may be directly obtained by referring to the related descriptions in the above method embodiments, which are not repeated herein.

As shown in fig. 12, the communication device 1200 includes a processor 1210. As an implementation method, the communication apparatus 1200 further includes an interface circuit 1220, and the processor 1210 and the interface circuit 1220 are coupled to each other. It is understood that the interface circuit 1220 may be a transceiver or an input-output interface. As an implementation method, the communication apparatus 1200 may further include a memory 1230, which is configured to store instructions executed by the processor 1210, or to store input data required for the processor 1210 to execute the instructions, or to store data generated after the processor 1210 executes the instructions.

When the communication device 1200 is used to implement the above-described method embodiment, the processor 1210 is used to implement the functions of the determining module 1110 and the training module 1120, and the interface circuit 1220 is used to implement the functions of the transceiver module 1130.

Illustratively, when the communications apparatus 1200 is configured to implement the function of the first network element in the above-described method embodiment, the processor 1210 is configured to invoke the computer program instructions to perform the following processing:

Determining a second network index according to the first network prediction model, wherein the input parameters of the first network prediction model comprise a first network index, the first network index comprises an actual network index of the network at a first moment and/or an actual network index before the first moment, the second network index is a network index of the network output by the first network prediction model according to the first network index at a second moment, and the second moment is later than the first moment;

Determining a first prediction deviation, wherein the first prediction deviation is the deviation of a second network index relative to a third network index, the third network index is an actual network index of the network at a second moment, and the network executes a first instruction between the first moment and the second moment, and the first instruction is used for adjusting network parameters of the network; according to the first prediction deviation and the first instruction, model parameters of the first network prediction model are adjusted, and a second network prediction model is obtained; when the input parameters of the second network prediction model comprise a first network index and a first instruction, the deviation of the network index of the network output by the second network prediction model at the second moment relative to the third network index is smaller than the first prediction deviation.

When the processor 1210 obtains the second network prediction model, the instructions may specifically be clustered according to the first instruction, or the instructions may be clustered according to the first network index and the second network index, to obtain a first class of instructions; wherein the first instruction belongs to a first class of instructions; and then, according to the first prediction deviation and the first type instruction, adjusting model parameters of the first network prediction model to obtain a second network prediction model. For example, the network indicator change amount corresponding to each instruction in the first class of instructions may be within a preset range.

The interface circuit 1220 may be used to: transmitting the first network index and/or the second network index to the second network element; and receiving a first instruction from the second network element, wherein the first instruction is output by a network control model running on the second network element, and input parameters of the network control model comprise a first network index and/or a second network index and further comprise a network running target.

Optionally, the network operation target includes a network operation target corresponding to the target user, and the first instruction includes a network adjustment parameter corresponding to the target user.

Optionally, the interface circuit 1220 is further configured to: and sending a second network index to the third network element, wherein the second network index is used for outputting the network index corresponding to the target user at the second moment by the user prediction model.

Optionally, the input parameters of the first network prediction model further include a first instruction, and the second network index is a network index of the network output by the first network prediction model at the second moment according to the first network index and the first instruction.

Illustratively, when the communications apparatus 1200 is configured to implement the function of the second network element in the method embodiment described above, the processor 1210 is configured to invoke the computer program instructions to perform the following processing:

Determining a first instruction according to a first network control model; the input parameters of the first network control model comprise a network operation target and at least one network index, the first instruction is an instruction which is output by the first network control model and is used for speculatively enabling the at least one network index to reach the network operation target, and the first instruction is used for adjusting the network parameters of a network;

Determining a first control deviation, wherein the first control deviation is the deviation of the change amount of at least one network index corresponding to the first instruction relative to the change amount of a target network index capable of achieving a network operation target; adjusting model parameters of the first network control model according to the first control deviation to obtain a second network control model; when the input parameters of the second network control model comprise a network operation target and at least one network index, the deviation of the change amount of at least one network index corresponding to the second instruction output by the second network control model relative to the change amount of the target network index is lower than the first control deviation.

Illustratively, the at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of at least one network index is the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index. The at least one network index is a second network index, wherein the second network index is a network index of the network at a second moment, and the second moment is later than the first moment, according to an actual network index of the network at the first moment and/or a network index of the network output by an actual network index before the first moment; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the second network index.

The at least one network index comprises a first network index and a second network index, wherein the first network index is an actual network index of the network at a first moment and/or an actual network index before the first moment, and the second network index is a network index of the network, output by the network prediction model according to the first network index, at a second moment, and the second moment is later than the first moment; the change amount of at least one network index is the change amount of a second network index after the network executes the first instruction or the change amount of a first network index after the network executes the first instruction; the target network index change amount is the difference value between the network operation target and the first network index.

Alternatively, the initial values of the model parameters of the first network control model may be determined based on the network intent.

Optionally, the interface circuit 1220 may be further configured to send a first instruction to the first network element, where the first instruction is configured to adjust a model parameter of a network prediction model running on the first network element.

Optionally, the network operation target may further include a network operation target corresponding to the target user, the first instruction includes a network adjustment parameter corresponding to the target user, and the input parameter of the first network control model further includes a network adjustment parameter corresponding to the target user; the interface circuit 1220 is configured to obtain, from the fourth network element, a network adjustment parameter corresponding to the target user, where the network adjustment parameter is output by a user control model running on the fourth network element.

It is to be appreciated that the processor in embodiments of the application may be a central processing unit (central processing unit, CPU), but may also be other general purpose processors, digital signal processors (DIGITAL SIGNAL processors, DSPs), application Specific Integrated Circuits (ASICs), field programmable gate arrays (field programmable GATE ARRAY, FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general purpose processor may be a microprocessor, but in the alternative, it may be any conventional processor.

It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims (26)

1. A model training method, characterized by comprising: The first network element determines a second network indicator according to a first network prediction model, wherein the input parameter of the first network prediction model includes the first network indicator, the first network indicator includes an actual network indicator of the network at a first moment and/or an actual network indicator before the first moment, the second network indicator is a network indicator of the network at a second moment output by the first network prediction model according to the first network indicator, and the second moment is later than the first moment; The first network element determines a first prediction deviation, where the first prediction deviation is a deviation of the second network indicator relative to a third network indicator, where the third network indicator is an actual network indicator of the network at the second moment, and the network executes a first instruction between the first moment and the second moment, where the first instruction is used to adjust a network parameter of the network; The first network element adjusts the model parameters of the first network prediction model according to the first prediction deviation and the first instruction to obtain a second network prediction model; wherein, when the input parameters of the second network prediction model include the first network indicator and the first instruction, the deviation of the network indicator of the network at the second moment output by the second network prediction model relative to the third network indicator is less than the first prediction deviation. 2. The method according to claim 1, wherein the first network element adjusts the model parameters of the first network prediction model according to the first prediction deviation and the first instruction to obtain the second network prediction model, comprising: The first network element clusters instructions according to the first instruction, or the first network element clusters instructions according to the first network indicator and the second network indicator to obtain first-category instructions; wherein the first instruction belongs to the first-category instructions; The first network element adjusts the model parameters of the first network prediction model according to the first prediction deviation and the first type of instructions to obtain a second network prediction model. 3. The method as claimed in claim 2 is characterized in that the network indicator change corresponding to each instruction in the first category of instructions is within a preset range. 4. The method according to any one of claims 1 to 3, characterized in that the method further comprises: The first network element sends the first network indicator and/or the second network indicator to the second network element; The first network element receives the first instruction from the second network element, wherein the first instruction is output by a network control model running on the second network element, and input parameters of the network control model include the first network indicator and/or the second network indicator, and also include network operation objectives. 5. The method according to claim 4, wherein the network operation target includes a network operation target corresponding to a target user, and the first instruction includes a network adjustment parameter corresponding to the target user. 6. The method according to claim 5, characterized in that the method further comprises: The first network element sends the second network indicator to the third network element, and the second network indicator is used for the user prediction model to output the network indicator corresponding to the target user at the second moment. 7. The method according to any one of claims 1-3 is characterized in that the input parameters of the first network prediction model also include the first instruction, and the second network indicator is the network indicator of the network at the second moment output by the first network prediction model based on the first network indicator and the first instruction. 8. A model training method, characterized by comprising: The second network element determines a first instruction according to the first network control model; wherein the input parameters of the first network control model include a network operation target and at least one network indicator, the first instruction is an instruction output by the first network control model that is inferred to make the at least one network indicator reach the network operation target, and the first instruction is used to adjust the network parameters of the network; The second network element determines a first control deviation, where the first control deviation is a deviation of a change amount of the at least one network indicator corresponding to the first instruction relative to a change amount of a target network indicator capable of achieving the network operation target; The second network element adjusts the model parameters of the first network control model according to the first control deviation to obtain a second network control model; wherein, when the input parameters of the second network control model include the network operation target and the at least one network indicator, the deviation of the change of the at least one network indicator corresponding to the second instruction output by the second network control model relative to the change of the target network indicator is lower than the first control deviation. 9. The method of claim 8, wherein the at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of the at least one network indicator is the change amount of the first network indicator after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the first network indicator; or, The at least one network indicator is a second network indicator, wherein the second network indicator is a network indicator of the network at a second moment output by a network prediction model based on the actual network indicator of the network at the first moment and/or the actual network indicator before the first moment, and the second moment is later than the first moment; the change amount of the at least one network indicator is the change amount of the second network indicator after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the second network indicator; or, The at least one network indicator includes a first network indicator and a second network indicator, wherein the first network indicator is an actual network indicator of the network at a first moment and/or an actual network indicator before the first moment, and the second network indicator is a network indicator of the network at a second moment output by a network prediction model based on the first network indicator, and the second moment is later than the first moment; the change in the at least one network indicator is the change in the second network indicator after the network executes the first instruction or the change in the first network indicator after the network executes the first instruction; the target network indicator change is the difference between the network operation target and the first network indicator. 10. The method according to claim 8 or 9, characterized in that the initial values of the model parameters of the first network control model are determined according to the network intent. 11. The method according to any one of claims 8 to 10, further comprising: The second network element sends the first instruction to the first network element, where the first instruction is used to adjust model parameters of the network prediction model running on the first network element. 12. The method according to any one of claims 8 to 11, characterized in that the network operation target includes a network operation target corresponding to a target user, the first instruction includes a network adjustment parameter corresponding to the target user, and the input parameter of the first network control model also includes a network adjustment parameter corresponding to the target user; The method also includes: the second network element obtains a network adjustment parameter corresponding to the target user from a fourth network element, and the network adjustment parameter is output by a user control model running on the fourth network element. 13. A communication device, comprising a processor and a memory: The memory stores computer program instructions; The processor is used to call the computer program instructions to perform the following processing: Determine a second network indicator according to a first network prediction model, wherein the input parameter of the first network prediction model includes a first network indicator, the first network indicator includes an actual network indicator of the network at a first moment and/or an actual network indicator before the first moment, the second network indicator is a network indicator of the network at a second moment output by the first network prediction model according to the first network indicator, and the second moment is later than the first moment; Determine a first prediction deviation, where the first prediction deviation is a deviation of the second network indicator relative to a third network indicator, where the third network indicator is an actual network indicator of the network at the second moment, and the network executes a first instruction between the first moment and the second moment, where the first instruction is used to adjust network parameters of the network; adjust model parameters of the first network prediction model according to the first prediction deviation and the first instruction to obtain a second network prediction model; wherein, when the input parameters of the second network prediction model include the first network indicator and the first instruction, the deviation of the network indicator of the network at the second moment output by the second network prediction model relative to the third network indicator is less than the first prediction deviation. 14. The device according to claim 13, wherein the processor is specifically configured to: Clustering instructions according to the first instructions, or clustering instructions according to the first network indicator and the second network indicator to obtain first-category instructions; wherein the first instruction belongs to the first-category instructions; The model parameters of the first network prediction model are adjusted according to the first prediction deviation and the first type of instructions to obtain a second network prediction model. 15. The device as claimed in claim 13, characterized in that the network indicator change amount corresponding to each instruction in the first category of instructions is within a preset range. 16. An apparatus as described in any one of claims 13-15, characterized in that the apparatus further comprises an interface circuit, wherein the interface circuit is used to send the first network indicator and/or the second network indicator to a second network element; and receive the first instruction from the second network element, wherein the first instruction is output by a network control model running on the second network element, and the input parameters of the network control model include the first network indicator and/or the second network indicator, and also include a network operation target. 17. The device according to claim 16, characterized in that the network operation target includes a network operation target corresponding to a target user, and the first instruction includes a network adjustment parameter corresponding to the target user. 18. The device according to claim 17, wherein the interface circuit is further used for: The second network indicator is sent to a third network element, where the second network indicator is used by a user prediction model to output a network indicator corresponding to the target user at the second moment. 19. The device according to any one of claims 13-15 is characterized in that the input parameters of the first network prediction model also include the first instruction, and the second network indicator is the network indicator of the network at the second moment output by the first network prediction model based on the first network indicator and the first instruction. 20. A communication device, comprising a processor and a memory: The memory stores computer program instructions; The processor is used to call the computer program instructions to perform the following processing: determine a first instruction according to a first network control model; wherein the input parameters of the first network control model include a network operation target and at least one network indicator, the first instruction is an instruction output by the first network control model that is inferred to make the at least one network indicator reach the network operation target, and the first instruction is used to adjust a network parameter of the network; Determine a first control deviation, where the first control deviation is a deviation of the change of the at least one network indicator corresponding to the first instruction relative to the target network indicator change that can achieve the network operation target; adjust the model parameters of the first network control model according to the first control deviation to obtain a second network control model; wherein, when the input parameters of the second network control model include the network operation target and the at least one network indicator, the deviation of the change of the at least one network indicator corresponding to the second instruction output by the second network control model relative to the target network indicator change is lower than the first control deviation. 21. The device according to claim 20, characterized in that the at least one network indicator is a first network indicator, wherein the first network indicator is an actual network indicator of the network; the change amount of the at least one network indicator is the change amount of the first network indicator after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the first network indicator; or, The at least one network indicator is a second network indicator, wherein the second network indicator is a network indicator of the network at a second moment output by a network prediction model based on the actual network indicator of the network at the first moment and/or the actual network indicator before the first moment, and the second moment is later than the first moment; the change amount of the at least one network indicator is the change amount of the second network indicator after the network executes the first instruction; the target network indicator change amount is the difference between the network operation target and the second network indicator; or, The at least one network indicator includes a first network indicator and a second network indicator, wherein the first network indicator is an actual network indicator of the network at a first moment and/or an actual network indicator before the first moment, and the second network indicator is a network indicator of the network at a second moment output by a network prediction model based on the first network indicator, and the second moment is later than the first moment; the change in the at least one network indicator is the change in the second network indicator after the network executes the first instruction or the change in the first network indicator after the network executes the first instruction; the target network indicator change is the difference between the network operation target and the first network indicator. 22. The device according to claim 20 or 21, characterized in that the initial values of the model parameters of the first network control model are determined according to the network intention. 23. The device according to any one of claims 20 to 22, further comprising: The interface circuit is used to send the first instruction to the first network element, where the first instruction is used to adjust the model parameters of the network prediction model running on the first network element. 24. The device according to any one of claims 20 to 23, characterized in that the network operation target includes a network operation target corresponding to a target user, the first instruction includes a network adjustment parameter corresponding to the target user, and the input parameter of the first network control model also includes a network adjustment parameter corresponding to the target user; The device also includes an interface circuit for acquiring a network adjustment parameter corresponding to the target user from a fourth network element, wherein the network adjustment parameter is output by a user control model running on the fourth network element. 25. A communication device, characterized by comprising a module for executing the method according to any one of claims 1 to 7 or a module for executing the method according to any one of claims 8 to 12. 26. A computer-readable storage medium, characterized in that a computer program or instruction is stored in the storage medium, and when the computer program or instruction is executed by a communication device, the method according to any one of claims 1 to 7 or the method according to any one of claims 8 to 12 is implemented.