CN113792538B - Method and device for rapidly generating operation ticket of power distribution network - Google Patents

Abstract

The invention provides a method and a device for rapidly generating operation tickets of a power distribution network. The method comprises the following steps: sequentially inputting a maintenance application form sentence text into a word segmentation model, a part-of-speech labeling model, a dependency syntax analysis model, a distribution network scheduling semantic role recognition model and a maintenance application form semantic recognition model, and outputting key information of operation tasks in different types of maintenance application forms; and converting the initial state and the termination state of the equipment in the key information into an environment initial state matrix and an environment termination state matrix, and then inputting the environment initial state matrix and the environment termination state matrix into a deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task. The invention can intelligently generate the dispatching instruction ticket, converts the artificial writing into automatic ticket forming of the system, releases the labor capacity of distribution network regulation personnel, and greatly saves ticket forming time.

Description

A method and device for quickly generating a distribution network operation ticket

Technical Field

The invention relates to a method and a device for quickly generating a distribution network operation ticket, and belongs to the field of distribution network regulation and management.

Background technique

An operation ticket is a set of operation instructions for switching the operating status of relevant equipment in a sequence described in standardized operation terms, in accordance with specific equipment and specific operation tasks, based on power grid safety operation procedures.

When issuing a dispatch operation ticket, the dispatcher first opens the maintenance application form, checks the corresponding equipment name and operating status, and proofreads the content of the maintenance application form; then analyzes the content of the maintenance application form and extracts the operation task; according to the operation task, searches for historical typical operation tickets in the dispatch management system, opens the relevant line map to analyze and check the line operation status, and manually infers the operation items based on the corresponding equipment name, topological relationship and historical typical ticket experience rules; finally, manually fills in the operation items one by one to generate an operation ticket.

As the scale of distribution networks continues to expand, the difficulty of regulation and management is also increasing. The distribution network is directly connected to a large number of electricity customers. How to improve the reliability of power supply and ensure that thousands of households have "no power outages or less power outages" has become the top priority for distribution network regulation and management personnel. The current wiring method of the distribution network is becoming more and more complicated, and the formulation of dispatching operation tickets is becoming more and more frequent. It is becoming increasingly difficult for dispatchers to control this increasingly complex distribution network system, and the problem of safe and stable operation of the distribution network is becoming increasingly severe. At present, the customization of operation ticket rules is complex, and the existing operation ticket management system cannot realize online real-time invoicing, and there is a lack of strict verification during the invoicing process. The invoicing work occupies a lot of time in the regulation business. It is necessary to shorten the time for generating maintenance operation tickets and reduce the workload of dispatching staff. A new method for generating operation tickets is proposed.

Summary of the invention

The purpose of the present invention is to provide a method for quickly generating a distribution network operation ticket, so as to solve the problems of cumbersome operation and long ticket generation time in the prior art manual ticket generation.

In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

In one aspect, the present invention provides a method for quickly generating a distribution network operation ticket, comprising:

Input the sentence text in the work content, maintenance ticket type, and power outage or power supply range columns in the maintenance application form into the conditional random field word segmentation model, and output the word list after word segmentation;

Input the segmented word list into a conditional random field part-of-speech tagging model, and output a word list with part-of-speech tags;

Input the word list marked with parts of speech into the dependency syntactic analysis model, and output a word list containing dependency relationships of each word;

Input the word list containing the dependency relationships of each word into the distribution network scheduling semantic role recognition model, and output a word list marked with the distribution network scheduling semantic role attributes;

Input the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and output the key information of the operation tasks in different types of maintenance application forms;

Converting the device initial state and final state in the key information into an environment initial state matrix and an environment final state matrix;

The environment initial state matrix and the environment terminal state matrix are input into a deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task.

Furthermore, the conditional random field word segmentation model is trained using a quasi-Newton algorithm, and the corpus used for training includes: Peking University word segmentation data, ICWB2 conference word segmentation data, MSR05 word segmentation data, and relevant corpus of power system distribution network scheduling.

Furthermore, the conditional random field part-of-speech tagging model is trained using the Viterbi algorithm, and the standard for part-of-speech tagging adopts Peking University's "Modern Chinese Corpus Processing Standards - Standards for Word Segmentation and Part-of-Speech Tagging", which uses a coarse-grained standard to divide parts of speech.

Furthermore, the dependency syntax analysis model is trained using a neural network algorithm, and the corpus used for training includes the semantic dependency network corpus tree library of Tsinghua University and relevant corpus of power system distribution network dispatching. Each sentence is annotated according to the CONLL format, and the annotation content includes serial number, word, prototype, coarse-grained part of speech, fine-grained part of speech, syntactic features, central word and dependency relationship.

Furthermore, the distribution network scheduling semantic role recognition model includes a bidirectional long short-term memory network and a conditional random field layer connected to the bidirectional long short-term memory network. The labeling method of the semantic role attributes is: one-hot encoding is performed according to the position of each word in the distribution network scheduling initial dictionary, and the code is used as the label of the distribution network scheduling semantic role.

Furthermore, the word list marked with the distribution network scheduling semantic role attribute is input into the maintenance application form semantic recognition model, and the key information of the operation tasks in different types of maintenance application forms is output, specifically including:

In the word list marked with the distribution network dispatching semantic role attribute, a synonym search is performed on the content of the operation type to determine the type of the maintenance application form;

According to the type of maintenance application form, the plant name, feeder name, equipment name involved in the operation, and equipment status name are extracted from the word list marked with the distribution network scheduling semantic role attribute; the prepositional object relationship in the dependency relationship is searched, and the equipment names involved in the operation are searched to obtain the equipment names of the starting and ending positions of the power outage section or the operation section, the equipment names of the new equipment started/exited, and the status names before and after the equipment operation; the parallel relationship in the dependency relationship is searched, and the sentences containing the parallel relationship are divided into multiple short sentences to obtain the key information of the maintenance application form operation task.

Furthermore, the converting of the device initial state and the terminal state in the key information into an environment initial state matrix and an environment terminal state matrix specifically includes:

According to the device ID in the key information, the ring network topology is retrieved, and the device initial state and terminal state in the key information are combined with the ring network topology to obtain an environment initial state matrix and an environment terminal state matrix.

Furthermore, the environment initial state matrix and the environment termination state matrix are input into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task, specifically including:

The initial state matrix of the environment is input into the deep learning reasoning network model, and the model outputs the state of the first step operation item. The changes of the input initial state matrix of the environment and the output state of the first step operation item are compared to obtain the first step operation item; the first step operation item state is continued to be input into the deep learning reasoning network model, and the second step operation item state is output. The changes of the input first step operation item state and the output second step operation item state are compared to obtain the second step operation item; and so on, until the operation item state output by the model is the same as the environment termination state matrix, and all ordered operation items corresponding to the operation task are obtained.

Furthermore, the aforementioned method for quickly generating distribution network operation tickets, after obtaining each operation item, also includes: performing anti-error verification on the obtained operation item; if the verification is successful, the final state of the operation item output by the model is continued to be input into the deep learning reasoning network model to output the final state of the operation item for the next step; if the verification is unsuccessful, the final state of the operation item for the previous step is restored and re-input into the deep learning reasoning network model for reasoning.

Furthermore, the deep learning reasoning network model is trained using a deep deterministic policy gradient algorithm, and the training method of the deep learning reasoning network model includes:

Get multiple historical operation tickets;

For each historical operation ticket, retrieve the ring network topology structure according to the device ID contained in it;

Based on the ring network topology, the initial state and final state of the equipment corresponding to each historical operation ticket are converted into the environment initial state matrix X ₀ and the environment final state matrix X _e ;

The initial state matrix _X0 of the environment corresponding to each historical operation ticket, the states _X1 , _X2 , ..., _XT of all operation items, where T is the number of operation items, are respectively input into the deep learning reasoning network model, and the model sequentially outputs variable 1, variable 2, ..., variable T. The deep learning reasoning network model is trained with the goal that variable 1 is the same as the state _X1 of the first step operation item, variable 2 is the same as the state _X2 of the second step operation item, ..., variable T is the same as the environment termination state matrix _Xe.

Furthermore, the conversion method of the environment state matrix includes:

For a single feeder tree network, assuming it contains n single-ended devices and m double-ended devices, the state of the single-ended device is represented by an n×1 matrix S:

S＝[x ₁ x ₂ x ₃ …x _n ] ^T

Where: x ₁ represents the status of the first single-ended device, and takes the value 0, 1, 2, or 3, where 0 represents "operation", 1 represents "hot standby", 2 represents "cold standby", and 3 represents "maintenance";

The state of the two-terminal device is represented by a 1×m matrix D:

D＝[e ₁ e ₂ e ₃ …e _m ]

Where, e ₁ represents the state of the first two-terminal device, and takes the value 0, 1, 2, or 3, where 0 represents "operation", 1 represents "hot standby", 2 represents "cold standby", and 3 represents "maintenance";

According to the ring network topology, the matrix S and the matrix D are associated to obtain the environmental state matrix K:

Where: the row represents a single-ended device, the column represents a two-ended device, k _1,1 represents the relationship between the first two-ended device and the first single-ended device, and k _1,1 takes the values of 1, 0, and -1, where 1 represents that the first single-ended device is the head end of the first two-ended device, -1 represents that the first single-ended device is the end end of the first two-ended device, and 0 represents no association.

In another aspect, the present invention provides a device for quickly generating a distribution network operation ticket, comprising:

The word segmentation module is configured to input the sentence text in the work content, maintenance ticket type, and power outage or power supply range column in the maintenance application form into the conditional random field word segmentation model, and output a word list after word segmentation;

A part-of-speech tagging module is configured to input the segmented word list into a conditional random field part-of-speech tagging model and output a word list tagged with part-of-speech;

A dependency syntactic analysis module, configured to input the word list marked with parts of speech into a dependency syntactic analysis model, and output a word list containing dependency relationships of each word;

A semantic role identification module, configured to input the word list containing the dependency relationship of each word into the distribution network scheduling semantic role identification model, and output a word list marked with the distribution network scheduling semantic role attribute;

A semantic recognition module, configured to input the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and output key information of the operation tasks in different types of maintenance application forms;

A preprocessing module, configured to convert the device initial state and the terminal state in the key information into an environment initial state matrix and an environment terminal state matrix;

The reasoning module is configured to input the environment initial state matrix and the environment terminal state matrix into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task.

The beneficial technical effects achieved by the present invention are as follows: (1) Natural language processing technology is used to perform semantic analysis on the text of maintenance application forms and task terms, thereby releasing the labor capacity of distribution network control personnel and realizing the association between operation tickets and application forms; (2) Dispatch instruction tickets are intelligently generated in combination with graphic model data, and manual drafting is converted into system-automated tickets, which greatly saves the work time of drafting tickets.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of a method for quickly generating a distribution network operation ticket according to an embodiment of the present invention;

Figure 2 is a sample diagram of a maintenance application form;

Figure 3 is a diagram showing an example of a topological structure.

Detailed ways

The present invention will be further described below in conjunction with specific examples. The following examples are only used to more clearly illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention.

As mentioned above, when the dispatcher issues a dispatch operation ticket, it is necessary to manually generate the operation ticket, but the manual ticket generation operation is cumbersome and takes a long time.

To this end, an embodiment of the present invention provides a method for quickly generating a distribution network operation ticket. As shown in FIG1 , the method includes the following steps:

Step 11, input the sentence text in the work content, maintenance ticket type, and power outage or power supply range columns in the maintenance application form into the conditional random field word segmentation model, and output the word list after word segmentation; first, use the quasi-Newton algorithm to train the conditional random field word segmentation model.

The word segmentation model uses common corpora, including: "Peking University word segmentation data", "ICWB2 conference word segmentation data", and "MSR05 (Microsoft Research) word segmentation data".

In addition, relevant corpus of power system distribution network dispatching is collected, such as the maintenance content in the historical maintenance application form of the pilot unit and the distribution network control operation manual and regulations of the local power company, and it is also used as corpus for training the word segmentation model.

The BMSE four-structure word position tagging method is used to perform Chinese word segmentation tagging on the general corpus and the relevant corpus of power system distribution network scheduling, that is, B represents the beginning of the word, E represents the end of the word, M represents the middle of the word, and S represents a separate word.

The annotated general corpus and the relevant corpus of power system distribution network dispatching are input into the word segmentation model for training, and the trained model file is saved.

Then, when using the trained word segmentation model for word segmentation, the maintenance application sentence text is input into the model and a word list after word segmentation is output.

Step 12, inputting the segmented word list into a conditional random field part-of-speech tagging model, and outputting a word list with part-of-speech tags;

Firstly, the Viterbi algorithm is used to train the conditional random field part-of-speech tagging model.

The content of the corpus for part-of-speech tagging is consistent with that of the corpus for word segmentation, including general corpus and corpus related to power system distribution network dispatching. At the same time, part-of-speech tagging needs to be added based on the word segmentation corpus.

The standard for part-of-speech tagging adopts Peking University's "Modern Chinese Corpus Processing Standards - Standards for Word Segmentation and Part-of-Speech Tagging", which uses a coarse-grained standard to divide parts of speech.

Input the annotated corpus into the part-of-speech tagging model for training and save the trained model file.

Then, when using the trained part-of-speech tagging model for part-of-speech tagging, the sentence of the maintenance application form after word segmentation is input into the part-of-speech tagging model, and a list of words with parts of speech is output.

Step 13, inputting the word list marked with parts of speech into the dependency syntactic analysis model, and outputting a word list containing dependency relationships of each word;

Firstly, a neural network algorithm is used to train the dependency parsing model.

The common corpus for dependency syntactic analysis is the semantic dependency network corpus tree library of Tsinghua University. At the same time, the sentence contents in the historical maintenance application forms of the pilot units and the materials related to the power system distribution network dispatching are collected.

Each sentence in the corpus is annotated according to the CONLL format. The annotation content includes serial number, word, prototype (Chinese word = prototype), coarse-grained part of speech, fine-grained part of speech, syntactic features, central word (the serial number of the central word in the sentence) and dependency relationship. Since the coarse-grained part of speech has been annotated during the part-of-speech tagging training process, the coarse-grained part of speech and fine-grained part of speech in the CONLL format are consistent with the part of speech of the corpus during the part-of-speech tagging training process.

The annotated general dependency sentence data and the dependency sentence data related to the power system distribution network dispatching are input into the dependency syntax analysis model for training, and the trained model file is saved.

After training, you can use the dependency syntactic analysis model to perform dependency syntactic analysis. Input the word list with parts of speech in the sentence into the dependency syntactic analysis model to obtain the dependency relationship of each word in the sentence. The output format is CONLL format.

Step 14, inputting the word list containing the dependency relationships of each word into a distribution network scheduling semantic role recognition model, and outputting a word list marked with distribution network scheduling semantic role attributes;

The distribution network scheduling semantic role recognition model adopts the BiLSTM-CRF model, which includes a bidirectional long short-term memory network (BiLSTM) layer and a conditional random field (CRF) layer connected to it. The BiLSTM layer predicts the label score of the semantic role of each word unit, and the CRF layer adds sequence constraints to the predicted labels to improve the accuracy of model recognition.

First, the BiLSTM-CRF model is trained.

Based on the corpus of the dependency syntax analysis in the CONLL format in step 13, add the annotation of the semantic role of the maintenance application form for each word. The semantic role includes: plant name, feeder name, equipment name involved in the operation, equipment status name, operation type (such as power outage, power supply) and other information.

In a specific embodiment, one-hot encoding is performed on each word according to its position in the initial scheduling dictionary, and the encoding serves as a label for the semantic role of the distribution network scheduling.

Input the annotated corpus into the BiLSTM-CRF model for training, and save the model file after training is completed.

After training, the BiLSTM-CRF model can be used to identify semantic roles for distribution network scheduling. A sentence in CONLL format is input, and the BiLSTM layer predicts the label score of each word semantic role. All scores predicted by the BiLSTM layer are input into the CRF layer. In the CRF layer, the highest label score among all semantic roles of each word unit is used as the semantic role of the word.

Step 15, inputting the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and outputting the key information of the operation tasks in different types of maintenance application forms;

After the maintenance application form sentences have been segmented, POS-tagged, parsed for dependency syntax, and identified for distribution network dispatching semantic roles, the key information is identified based on the output results of the above analysis according to different maintenance application form categories, and finally the semantic parsing of the maintenance application form is completed.

First, we searched for synonyms in the operation type column of the maintenance application form and divided the maintenance application forms into four types: "power outage", "power supply", "mode adjustment", and "new equipment startup/shutdown". Then, based on the above analysis results, we extracted key information from the four types of maintenance application statements:

a) The key information that needs to be extracted from the power outage maintenance application form includes: plant name, feeder name, operation type, equipment name at the start position of the power outage section, equipment name at the end position of the power outage section, starting state before operation, and ending state after operation.

b) The key information that needs to be extracted from the power transmission maintenance application form includes: plant name, feeder name, operation type, equipment name at the starting position of the power transmission section, equipment name at the ending position of the power transmission section, starting status before operation, and ending status after operation.

c) The key information that needs to be extracted from the mode adjustment maintenance application form includes: plant name, feeder name, mode adjustment type, equipment name at the starting position of the operation interval, equipment name at the ending position of the operation interval, starting state before operation, and ending state after operation.

d) The key information that needs to be extracted from the new equipment startup/shutdown maintenance application form includes: plant name, feeder name, startup/shutdown equipment name, and operation type.

After the above steps 11-14, each maintenance application form sentence has completed word segmentation, and each word contains the results of part-of-speech tagging, dependency syntax analysis, and maintenance application form semantic role tagging. The sentence containing these results serves as the input for maintenance application form semantic recognition.

Therefore, this step may specifically include:

Step 151, performing a synonym search on the operation type in the result of semantic role labeling to determine the type of the maintenance application form;

Step 152: extract key information of the maintenance application form operation task according to the type of the maintenance application form.

First, from the results of semantic role labeling, extract information such as plant name, feeder name, device name involved in the operation, and device status name;

Then, search for the preposition-object relationship in the dependency relationship, and search the names of the equipment involved in the operation to obtain the names of the starting and ending equipment of the power outage interval or the operation interval, the names of the equipment started/exited by the new equipment, and the names of the status of the equipment before and after the operation;

Finally, the parallel relationship in the dependency relationship is searched, and the sentences containing the parallel relationship are divided into multiple short sentences to complete the extraction of key information of the maintenance application sentence.

When the maintenance application form procedures of the pilot unit change, the key information extraction rules can be modified according to the actual situation.

The final output of the semantic parsing of the maintenance application form is a list of key information in the sentence, such as {"plant station name": Dugang Station, "feeder line name": Zhuanggang Line, "operation type": maintenance, "operation interval starting position equipment name": #47 pole drop-out fuse, "operation interval ending position equipment name": terminal line, "starting state before operation": operation, "ending state after operation": maintenance}.

In addition, the information in the maintenance application form other than the "work content" is in a structured table format and can be directly extracted according to the actual situation of the pilot unit, such as: transfer order type, operation purpose, start time, completion time, maintenance ticket application form ID and other information.

Step 16: Convert the device initial state and terminal state in the key information into an environment initial state matrix and an environment terminal state matrix.

In one embodiment, according to the device ID in the key information, the ring network topology is retrieved, and the device initial state and terminal state in the key information are combined with the ring network topology to obtain the environment initial state matrix and the environment terminal state matrix. For an electric power network, the electrical devices such as single-end devices (switches) and double-end devices (lines and transformers) are connected into a connection diagram, and each device has a different state. Each operation item in the operation ticket corresponds to the change of the state of the electrical device.

The electrical operating regulations stipulate that electrical equipment has four basic working states: operation, hot standby, cold standby and maintenance state:

(1) Operating status: The circuit breaker corresponding to the equipment and the switches on both sides should be in the closed state. In this state, the equipment has a certain voltage;

(2) Hot standby state: The circuit breaker of the equipment will be in the open position, and the knife switches on both sides will be in the closed state. As long as the circuit breaker is closed, the equipment will immediately switch to the operating state;

(3) Cold standby state: the circuit breaker and the switches on both sides are in the open position;

(4) Maintenance status: The circuit breaker and the switches on both sides of the equipment must be in the open position, and the grounding switch must be closed.

On this basis, the data is preprocessed to process the historical operation tickets and topological structure information into the environment state matrix of the deep learning inference network model.

Specifically, in a single feeder tree network, data preprocessing includes the following processes:

(1) Single-ended device - switch

Assuming the total number of switches in the topology is n, the switch state is represented by an n×1 matrix S:

S＝[x ₁ x ₂ x ₃ …x _n ] ^T

Where: _x1 represents the state of the first switch, which can take values 0, 1, 2 or 3, 0 represents "running", 1 represents "hot standby", 2 represents "cold standby" and 3 represents "maintenance".

(2) Two-terminal equipment - lines and transformers

Assuming the total number of two-terminal devices is m, the two-terminal device state is represented by a 1×m matrix D:

D＝[e ₁ e ₂ e ₃ …e _m ]

Where, _e1 represents the state of the first two-terminal device, and takes values of 0, 1, 2, or 3, where 0 represents "operation", 1 represents "hot standby", 2 represents "cold standby", and 3 represents "maintenance". When the line considers whether FA is put into use, 0 represents "operation + FA is put into use", 1 represents "hot standby", 2 represents "cold standby", 3 represents "maintenance", and 4 represents "operation + FA is disabled".

(3) Topological relationship

The topological relationship is represented by an n×m directed incidence matrix K:

In the formula: rows represent switches, columns represent lines or transformers, k _1,1 represents the relationship between the first line or transformer and the first switch, and k _1,1 takes values of 1, 0, and -1, where 1 represents that the switch is the beginning of the line or transformer, -1 represents that the switch is the end of the line or transformer, and 0 represents no association.

For example, for the topological structure shown in Figure 3, its topological matrix can be expressed as:

(4) Environmental state matrix X

The environment state matrix X of the deep learning inference network model contains both the state and topological relationship of the device, which can be expressed as:

Step 17: Input the environment initial state matrix and the environment terminal state matrix into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task.

In this step, the deep learning reasoning network model uses the Deep Deterministic Policy Gradient (DDPG) algorithm to perform intelligent reasoning on the operation items.

Intelligent reasoning of operation items is based on the operation ticket drafting rules, and according to the key information of the operation task and the power grid topology obtained by semantic analysis, the orderly operation items corresponding to the current operation task are inferred.

Before using the DDPG deep learning inference network model for inference, the model needs to be trained first.

Before model training, it is necessary to first preprocess the historical operation tickets and process the historical operation tickets and topological structure information into the environment state matrix of the DDPG deep learning inference network model as the input of the model training. The specific processing process is described in step 16.

The training of the DDPG deep learning inference network model is carried out with each operation ticket as one training round. The process is as follows:

① Enter a historical operation ticket, call the "Topology Analysis" service according to the device ID, and obtain the ring network topology structure;

② Combined with the ring network topology, the initial state and final state of the electrical equipment in the operation task are converted to obtain the initial matrix _X0 of the environmental state and the final state matrix _Xe .

③ Assume that there are T operation items in the operation ticket, then the state of the first operation item is _X1 , ..., the state of the Tth operation item is _XT , and _X0 , _X1 , ..., _XT , _Xe are input into the inference model in sequence to complete a round of training.

It should be noted that the operation item refers to the change from _X0 to _X1 , from _X1 to _X2 , ..., _XT-1 to _XT ; the operation item state refers to the final state of the operation item (ie, change), ie, _X1 , ..., _XT .

After training with a large number of historical operation tickets, save the model file.

After training, the DDPG deep learning inference network model is used to infer the key information in the operation task obtained through semantic analysis in step 15, thereby generating an operation ticket containing ordered operation items. The specific process is as follows:

① The key information of the operation task obtained by semantic analysis of the maintenance application form includes the device ID, the device start and end status, etc.; call the "topology analysis" service according to the device ID to obtain the ring network topology structure, process the device start and end status, and obtain the environment state initial matrix _X0 and the end state matrix _Xe .

② Input _X0 and _Xe into the DDPG deep learning reasoning network model, and the model reasoning is performed step by step.

③ The reasoning is carried out one step, and the intermediate variable _Xt is output. By comparing the changes in Xt _-1 and _{Xt ,} the content of the operation item can be obtained. For example, if the state of a switch changes from 0 to 1, the corresponding operation item is: "Change a switch from operation to hot standby."

④ Compare _Xt with the terminal state _Xe . If _Xt = _Xe , the reasoning ends. If _Xt and _Xe are inconsistent, return to step ③.

In a further implementation, the "anti-error verification" service can be called to perform the "five-prevention" verification of the inferred operation item _At . If the verification succeeds, the next inference is continued. If the verification fails, the state Xt _-1 is restored and the reasoning is repeated in step ③.

In addition, during the reasoning process, specified operation items can be added in combination with the special requirements of the trial units and cities.

Through the above embodiments, a method for quickly generating distribution network operation tickets of the present invention parses the content of the maintenance application form through semantic recognition technology, extracts key information such as the maintenance power outage scope, the status before and after the operation, the operation type, etc., generates operation items based on intelligent reasoning of the expert knowledge base, and sorts out the expert knowledge base rules to generate a reasoning model through training, and automatically generates operation items for the operation tasks extracted from the maintenance application form. The operation is simple, saves manual labor, and shortens the ticketing time.

Based on the above method, the operation process can be simulated and rehearsed through a graphical human-computer interface.

Specifically:

Based on the graphical human-machine interface, the system automatically performs operations in sequence according to the contents of the operation ticket, automatically changes the state of the operation object, and automatically checks whether each operation item is correct through simulated operation.

Initialize the preview line status according to the initial status of the operation ticket preview.

The system automatically loads the operation instructions of the operation ticket and performs simulation preview one by one. By changing the state of the operation object, it can truly reflect the actual process of the operation ticket execution; and provides the following functions such as setting breakpoints, single-step preview, automatic preview, and recovery:

Automatic preview: The system automatically executes all the operation ticket instructions in the operation ticket according to the operation sequence, and displays the status of each operation object after execution on the human-machine interface.

Set breakpoints: Set a power outage in a certain operation instruction in the operation ticket. During automatic preview, if there is a breakpoint, the system will automatically execute all instructions before the breakpoint in the operation sequence and display the current status of each operation object on the human-machine interface.

Single-step preview: Based on the previous step operation instruction, execute one step of operation instruction in sequence, and display the status of each operation object after execution on the human-machine interface.

Restore: Restore to the state of each operation object before the operation instruction in the operation ticket is executed.

After the preview is finished, you can repeat the preview, request to restore the initial state, initialize the preview line, and then perform simulation preview line by line.

The graphic-based operation ticket simulation rehearsal can simulate the execution of the contents of the operation ticket in sequence in the graphic, intuitively display the operation status of the power grid after the operation, reduce the dispatcher's misjudgment and erroneous execution, and effectively improve the level of safe dispatching operation.

In another embodiment, the present invention provides a device for quickly generating a distribution network operation ticket, comprising:

The word segmentation module is configured to input the sentence text in the work content, maintenance ticket type, and power outage or power supply range column in the maintenance application form into the conditional random field word segmentation model, and output a word list after word segmentation;

A part-of-speech tagging module is configured to input the segmented word list into a conditional random field part-of-speech tagging model and output a word list tagged with part-of-speech;

A dependency syntactic analysis module, configured to input the word list marked with parts of speech into a dependency syntactic analysis model, and output a word list containing dependency relationships of each word;

A semantic role identification module, configured to input the word list containing the dependency relationship of each word into the distribution network scheduling semantic role identification model, and output a word list marked with the distribution network scheduling semantic role attribute;

A semantic recognition module, configured to input the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and output key information of the operation tasks in different types of maintenance application forms;

A preprocessing module, configured to convert the device initial state and the terminal state in the key information into an environment initial state matrix and an environment terminal state matrix;

The reasoning module is configured to input the environment initial state matrix and the environment terminal state matrix into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task.

Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

The present application is described with reference to the flowchart and/or block diagram of the method, device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and/or box in the flowchart and/or block diagram, and the combination of the process and/or box in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one process or multiple processes in the flowchart and/or one box or multiple boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims (7)

1. A method for quickly generating a distribution network operation ticket, characterized by comprising: Input the sentence text in the work content, maintenance ticket type, and power outage or power supply range columns in the maintenance application form into the conditional random field word segmentation model, and output the word list after word segmentation; Input the segmented word list into a conditional random field part-of-speech tagging model, and output a word list with part-of-speech tags; Input the word list marked with parts of speech into the dependency syntactic analysis model, and output a word list containing dependency relationships of each word; Input the word list containing the dependency relationships of each word into the distribution network scheduling semantic role recognition model, and output a word list marked with the distribution network scheduling semantic role attributes; Input the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and output the key information of the operation tasks in different types of maintenance application forms; Converting the device initial state and final state in the key information into an environment initial state matrix and an environment final state matrix; Inputting the environment initial state matrix and the environment terminal state matrix into a deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task; The word list marked with the distribution network scheduling semantic role attribute is input into the maintenance application form semantic recognition model, and the key information of the operation tasks in different types of maintenance application forms is output, specifically including: In the word list marked with the distribution network dispatching semantic role attribute, a synonym search is performed on the content of the operation type to determine the type of the maintenance application form; According to the type of maintenance application form, extract the plant name, feeder name, equipment name involved in the operation, and equipment status name from the word list marked with the semantic role attribute of distribution network scheduling; search for the prepositional object relationship in the dependency relationship, and search for the equipment name involved in the operation to obtain the equipment name of the starting and ending position of the power outage section or operation section, the equipment name of the new equipment startup/exit, and the status name before and after the equipment operation; search for the parallel relationship in the dependency relationship, divide the sentence containing the parallel relationship into multiple short sentences, and obtain the key information of the maintenance application form operation task; The inputting the environment initial state matrix and the environment terminal state matrix into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task specifically includes: The initial state matrix of the environment is input into the deep learning reasoning network model, and the model outputs the state of the first step operation item. The change of the input initial state matrix of the environment and the output state of the first step operation item is compared to obtain the first step operation item; the state of the first step operation item is continuously input into the deep learning reasoning network model, and the state of the second step operation item is output to obtain the second step operation item. The change of the input state of the first step operation item and the output state of the second step operation item is compared to obtain the second step operation item; and so on, until the state of the operation item output by the model is the same as the environment termination state matrix, and all ordered operation items corresponding to the operation task are obtained; The deep learning reasoning network model is trained using a deep deterministic policy gradient algorithm, and the training method of the deep learning reasoning network model includes: Get multiple historical operation tickets; For each historical operation ticket, retrieve the ring network topology structure according to the device ID contained in it; Based on the ring network topology, the initial state and final state of the equipment corresponding to each historical operation ticket are converted into the environment initial state matrix X ₀ and the environment final state matrix X _e ; The initial state matrix _X0 of the environment corresponding to each historical operation ticket, the states _X1 , _X2 , ..., _XT of all operation items, where T is the number of operation items, are respectively input into the deep learning reasoning network model, and the model sequentially outputs variable 1, variable 2, ..., variable T. The deep learning reasoning network model is trained with the goal that variable 1 is the same as the state _X1 of the first step operation item, variable 2 is the same as the state _X2 of the second step operation item, ..., variable T is the same as the environment termination state matrix _Xe. 2. A method for quickly generating distribution network operation tickets according to claim 1, characterized in that the conditional random field word segmentation model is trained using a quasi-Newton algorithm, and the corpus used for training includes: Peking University word segmentation data, ICWB2 conference word segmentation data, MSR05 word segmentation data, and relevant corpus of power system distribution network scheduling. 3. A method for quickly generating distribution network operation tickets according to claim 1, characterized in that the conditional random field part-of-speech tagging model is trained using the Viterbi algorithm, and the standard for part-of-speech tagging adopts Peking University's "Modern Chinese Corpus Processing Specifications - Standards for Word Segmentation and Part-of-Speech Tagging", and uses a coarse-grained standard to divide the parts of speech. 4. According to claim 1, a method for quickly generating a distribution network operation ticket is characterized in that the dependency syntactic analysis model is trained using a neural network algorithm, and the corpus used for training includes the semantic dependency network corpus tree library of Tsinghua University and relevant corpus of power system distribution network scheduling. Each sentence is annotated according to the CONLL format, and the annotation content includes serial number, word, prototype, coarse-grained part of speech, fine-grained part of speech, syntactic features, central word and dependency relationship. 5. According to a method for quickly generating distribution network operation tickets according to claim 1, it is characterized in that the distribution network scheduling semantic role recognition model includes a bidirectional long short-term memory network and a conditional random field layer connected to the bidirectional long short-term memory network, and the annotation method of the semantic role attributes is: one-hot encoding is performed according to the position of each word in the initial dictionary of distribution network scheduling, and the encoding is used as the annotation of the distribution network scheduling semantic role. 6. A method for quickly generating a distribution network operation ticket according to claim 1, characterized in that the step of converting the equipment initial state and terminal state in the key information into an environment initial state matrix and an environment terminal state matrix specifically comprises: According to the device ID in the key information, the ring network topology is retrieved, and the device initial state and terminal state in the key information are combined with the ring network topology to obtain an environment initial state matrix and an environment terminal state matrix. 7. A device for quickly generating a distribution network operation ticket, characterized by comprising: The word segmentation module is configured to input the sentence text in the work content, maintenance ticket type, and power outage or power supply range column in the maintenance application form into the conditional random field word segmentation model, and output a word list after word segmentation; A part-of-speech tagging module is configured to input the segmented word list into a conditional random field part-of-speech tagging model and output a word list tagged with part-of-speech; A dependency syntactic analysis module, configured to input the word list marked with parts of speech into a dependency syntactic analysis model, and output a word list containing dependency relationships of each word; A semantic role identification module, configured to input the word list containing the dependency relationship of each word into the distribution network scheduling semantic role identification model, and output a word list marked with the distribution network scheduling semantic role attribute; A semantic recognition module, configured to input the word list marked with the distribution network scheduling semantic role attribute into the maintenance application form semantic recognition model, and output key information of the operation tasks in different types of maintenance application forms; A preprocessing module, configured to convert the device initial state and the terminal state in the key information into an environment initial state matrix and an environment terminal state matrix; A reasoning module is configured to input the environment initial state matrix and the environment terminal state matrix into a deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task; The word list marked with the distribution network scheduling semantic role attribute is input into the maintenance application form semantic recognition model, and the key information of the operation tasks in different types of maintenance application forms is output, specifically including: In the word list marked with the distribution network dispatching semantic role attribute, a synonym search is performed on the content of the operation type to determine the type of the maintenance application form; According to the type of maintenance application form, extract the plant name, feeder name, equipment name involved in the operation, and equipment status name from the word list marked with the semantic role attribute of distribution network scheduling; search for the prepositional object relationship in the dependency relationship, and search for the equipment name involved in the operation to obtain the equipment name of the starting and ending position of the power outage section or operation section, the equipment name of the new equipment startup/exit, and the status name before and after the equipment operation; search for the parallel relationship in the dependency relationship, divide the sentence containing the parallel relationship into multiple short sentences, and obtain the key information of the maintenance application form operation task; The inputting the environment initial state matrix and the environment terminal state matrix into the deep learning reasoning network model for step-by-step reasoning to obtain all ordered operation items corresponding to the operation task specifically includes: The initial state matrix of the environment is input into the deep learning reasoning network model, and the model outputs the state of the first step operation item. The change of the input initial state matrix of the environment and the output state of the first step operation item is compared to obtain the first step operation item; the state of the first step operation item is continuously input into the deep learning reasoning network model, and the state of the second step operation item is output to obtain the second step operation item. The change of the input state of the first step operation item and the output state of the second step operation item is compared to obtain the second step operation item; and so on, until the state of the operation item output by the model is the same as the environment termination state matrix, and all ordered operation items corresponding to the operation task are obtained; The deep learning reasoning network model is trained using a deep deterministic policy gradient algorithm, and the training method of the deep learning reasoning network model includes: Get multiple historical operation tickets; For each historical operation ticket, retrieve the ring network topology structure according to the device ID contained in it; Based on the ring network topology, the initial state and final state of the equipment corresponding to each historical operation ticket are converted into the environment initial state matrix X ₀ and the environment final state matrix X _e ; The initial state matrix _X0 of the environment corresponding to each historical operation ticket, the states _X1 , _X2 , ..., _XT of all operation items, where T is the number of operation items, are respectively input into the deep learning reasoning network model, and the model sequentially outputs variable 1, variable 2, ..., variable T. The deep learning reasoning network model is trained with the goal that variable 1 is the same as the state _X1 of the first step operation item, variable 2 is the same as the state _X2 of the second step operation item, ..., variable T is the same as the environment termination state matrix _Xe.