CN118504991B

CN118504991B - A method, device, equipment and medium for processing power outages in power security areas

Info

Publication number: CN118504991B
Application number: CN202410961992.8A
Authority: CN
Inventors: 韩荣杰; 来益博; 唐铁英; 黄江宁; 郑伟彦; 黄迪; 陈潘霞
Original assignee: Zhejiang Dayou Industrial Co ltd Hangzhou Science And Technology Development Branch; Hangzhou Power Supply Co of State Grid Zhejiang Electric Power Co Ltd
Current assignee: Zhejiang Dayou Industrial Co ltd Hangzhou Science And Technology Development Branch; Hangzhou Power Supply Co of State Grid Zhejiang Electric Power Co Ltd
Priority date: 2024-07-18
Filing date: 2024-07-18
Publication date: 2025-01-21
Anticipated expiration: 2044-07-18
Also published as: CN118504991A

Abstract

The invention discloses a power outage processing method, device, equipment and medium for a power guarantee area, wherein the method comprises the steps of marking time at the moment when a power outage occurs according to a high-precision time synchronization method, collecting historical operation data of all nodes of a power system at the same time point, extracting features of the operation data based on a deep learning algorithm, extracting depth feature vectors reflecting the intrinsic structure and time scale features of the data, establishing a power outage risk assessment model according to the depth feature vectors, optimizing the power outage risk assessment model by adopting a Bayesian network or Markov decision process method, obtaining current operation data of the power system, inputting the current operation data into the optimized power outage risk assessment model, outputting an assessment result, and constructing a power outage processing strategy matched with the assessment result and executed by a safety early warning terminal. The method provided by the invention realizes accurate monitoring and effective control of the power failure risk of the power system.

Description

Power failure processing method, device, equipment and medium for power guarantee area

Technical Field

The present invention relates to the field of information technologies, and in particular, to a power outage processing method, apparatus, device, and medium for a power protection area.

Background

In the power outage processing, a technical contradiction exists between the accuracy of a time scale and the accuracy of outage risk judgment. On the one hand, improving the accuracy of the time scale is helpful for more accurately evaluating the power failure risk, and particularly when distinguishing short-frequency power failure from long-time single power failure, the former is long and short in single power failure, but the frequent occurrence of the characteristics can bring about larger influence and inconvenience. On the other hand, improving the time scale accuracy means that more complex system and algorithm support is required, which may lead to an increase in system complexity, thereby increasing the probability of system failure and erroneous judgment.

In an actual power outage handling service, a trade-off between time accuracy and system complexity is required. It is found that the blackout risk level under different time accuracies can be quantified by constructing a risk function model taking time control parameters as variables. The model can intuitively display the relation between time parameter adjustment and risk change, and provides a basis for decision making. However, in the process of constructing the risk function model, technical challenges are faced in how to select proper time control parameters to accurately reflect the time characteristics of the outage event, secondly, the form of the risk function needs to be carefully designed, the influence of time precision is considered, other factors such as outage frequency and influence range are considered, and thirdly, the calculation efficiency of the model is a key problem, particularly in the case of large-scale power grid and massive outage data, how to improve the calculation speed and reduce resource consumption while ensuring the accuracy of the model.

Disclosure of Invention

The invention provides a power outage processing method, device, equipment and storage medium for a power outage guarantee area, which are used for solving the technical problem of balance between time precision and system complexity in actual power outage processing business, realizing accurate monitoring and effective control of power outage risk of a power system and improving the safety and stability of the power system.

In order to solve the above technical problems, an embodiment of the present invention provides a power outage processing method in a power protection area, including:

performing time marking at the moment of power failure according to a high-precision time synchronization method, and collecting historical operation data of each node of the power system at the same time point;

Performing feature extraction and representation learning on the historical operation data based on a deep learning algorithm of a convolutional neural network, automatically learning a local mode and a global mode in the historical data through a convolutional layer and a pooling layer, and extracting a deep feature vector reflecting the intrinsic structure and time scale features of the data;

Establishing a power failure risk assessment model according to the depth feature vector, and optimizing the power failure risk assessment model by adopting a Bayesian network or Markov decision process method;

and acquiring current operation data of the power system, inputting the current operation data into an optimized power outage risk assessment model, outputting an assessment result, and constructing a power outage processing strategy matched with the assessment result and executed by a safety early warning terminal.

As one of the preferable schemes, a time synchronization technology is combined with power failure detection, the voltage waveform of the power grid is monitored in real time through a high-speed voltage sampling and edge detection algorithm, and the accurate moment of power failure is recorded.

As one preferable mode, after collecting the operation data of each node of the power system at the same time point, the method further comprises:

Acquiring a synchronous time stamp according to the high-precision time synchronization method, performing time calibration and alignment on the acquired power outage event data, obtaining accurate power outage duration by calculating the start-stop time difference of the power outage event, and performing aggregation statistics according to different time scales to form a power outage duration distribution diagram with multiple granularities;

and (3) carrying out spatial correlation on the power outage event data and a power grid topological structure by adopting a geographic information system technology, positioning the influence range of the power outage event by sequentially passing through a buffer area analysis technology and a spatial connection algorithm, and calculating the number of affected users and the power load to form a visual thermodynamic diagram of the influence range of the power outage.

As one of the preferred schemes, after extracting the depth feature vector reflecting the internal structure and time scale features of the data, the method further comprises:

Inputting the processed depth feature vector into a multi-layer perceptron MLP network for training to obtain an anomaly identification model;

Establishing a historical experience database based on a graph database, and acquiring various abnormal mode types and corresponding treatment schemes which occur in the running process of a storage system;

Acquiring current operation data, inputting the current operation data into the anomaly identification model, carrying out matching processing on the new anomaly mode based on a graph matching algorithm after identifying the new anomaly mode, triggering a dynamic misjudgment risk assessment mechanism if matching fails, and otherwise, directly calling a corresponding treatment scheme.

As one preferable solution, the establishing a blackout risk assessment model according to the depth feature vector, and optimizing the blackout risk assessment model by adopting a bayesian network or markov decision process method includes:

Acquiring time sequence data in the running process of the power grid in real time through high-speed data acquisition equipment and an intelligent sensor;

Processing the time sequence data based on a signal processing technology, and extracting a power failure risk assessment index;

processing the blackout risk assessment index by adopting a data mining and machine learning algorithm to construct a blackout risk assessment model;

And according to the blackout risk assessment model, adopting a Bayesian network reasoning algorithm to quantitatively calculate probability parameters of each risk factor output by the blackout risk assessment model, and dynamically updating the probability parameters through maximum likelihood estimation and an expected maximization algorithm so as to realize real-time reasoning and prediction of blackout risk.

As one preferable solution, the constructing a matched outage processing policy executed by the safety early warning terminal according to the output result of the optimized outage risk assessment model includes:

Processing a power failure risk assessment result output by the power failure risk assessment model by adopting an equidistant dividing method to obtain a corresponding risk level;

and constructing a power failure processing strategy executed by the safety early warning terminal based on the risk level.

As one preferable scheme, the safety precaution terminal comprises at least one of the following:

an operation and maintenance team terminal, an maintainer terminal or an electric power responsibility person terminal.

Another embodiment of the present invention provides a power outage processing apparatus for a power protection area, including:

the acquisition module is used for carrying out time marking at the moment of power failure according to the high-precision time synchronization method and acquiring historical operation data of each node of the power system at the same time point;

the extraction module is used for carrying out feature extraction and representation learning on the historical operation data based on a deep learning algorithm of the convolutional neural network, automatically learning a local mode and a global mode in the historical operation data through a convolutional layer and a pooling layer, and extracting depth feature vectors reflecting the intrinsic structure and time scale features of the data;

the evaluation module is used for establishing a power failure risk evaluation model according to the depth feature vector, and optimizing the power failure risk evaluation model by adopting a Bayesian network or Markov decision process method;

The output module is used for acquiring current operation data of the power system, inputting the current operation data into the optimized power outage risk assessment model, outputting an assessment result, and constructing a power outage processing strategy matched with the assessment result and executed by the safety early warning terminal.

Still another embodiment of the present invention provides a power outage handling apparatus of a power outage handling area, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, the processor implementing the power outage handling method of the power outage handling area as described above when executing the computer program.

Still another embodiment of the present invention provides a computer readable storage medium storing a computer program, where when the computer program is executed by a device in which the computer readable storage medium is located, a power outage processing method in a power protection area as described above is implemented.

Compared with the prior art, the embodiment of the invention has the beneficial effects that at least one of the following points is adopted:

(1) By installing the high-performance data acquisition system, the state of the system is monitored in real time, operation data such as key indexes of voltage, current, frequency and the like are automatically acquired, abnormal modes in the operation data are identified, the operation condition and potential risk of the system are evaluated in real time, and an algorithm strategy is optimized to reduce the influence on the system performance.

(2) And comparing the data pattern identified by the algorithm with patterns in a historical experience database, and automatically adjusting time scale parameters or prompting operation and maintenance personnel to perform manual inspection, so that the accuracy of risk assessment is improved.

(3) By making a power failure risk function model, the power failure risk under different time scales can be quantitatively analyzed, the model can be dynamically adjusted along with time, and continuous monitoring of a power system is realized.

Drawings

FIG. 1 is a flow chart of a power outage handling method in a power assurance area according to one embodiment of the present invention;

FIG. 2 is a schematic diagram of a power outage handling device in a power management area according to one embodiment of the present invention;

Fig. 3 is a schematic view of a power outage handling apparatus in a power assurance area according to one embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the invention, and the purpose of these embodiments is to provide a more thorough and complete disclosure of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In the description of the present application, the terms "first," "second," "third," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first", "a second", "a third", etc. may explicitly or implicitly include one or more such feature. In the description of the present application, unless otherwise indicated, the meaning of "a plurality" is two or more.

In the description of the present application, unless explicitly stated or limited otherwise, the terms "mounted," "connected," and "connected" are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected, mechanically connected, electrically connected, directly connected, indirectly connected via an intervening medium, or in communication between two elements. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and the like are used herein for descriptive purposes only and not to indicate or imply that the apparatus or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore should not be construed as limiting the application. The term "and/or" as used herein includes any and all combinations of one or more of the associated listed items. The specific meaning of the above terms in the present application will be understood in specific cases by those of ordinary skill in the art.

In the description of the present application, it should be noted that all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs unless defined otherwise. The terminology used in the description of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application, as the particular meaning of the terms described above in the present application will be understood to those of ordinary skill in the art in the detailed description of the application.

An embodiment of the present invention provides a power outage processing method for a power protection area, specifically, referring to fig. 1, fig. 1 shows a flow chart of the power outage processing method for the power protection area provided in one embodiment of the present invention, which includes steps S1 to S4:

s1, performing time marking at the moment of power failure according to a high-precision time synchronization method, and collecting historical operation data of each node of a power system at the same time point;

In the embodiment, a high-speed data acquisition card, such as a data acquisition card of a PCI-Express bus architecture, is provided with a multi-channel synchronous ADC (analog-to-digital converter), realizes synchronous sampling of a plurality of measuring points, has a sampling rate of thousands of Hz to tens of thousands of Hz, and ensures real-time acquisition and low-delay transmission of data by matching with a cache and DMA transmission technology.

And installing intelligent sensors at each power system node, including a voltage transformer, a current transformer, a frequency measurement unit and the like, converting analog quantity into digital quantity, and uploading acquired data to a data center in real time through industrial Ethernet or optical fiber communication. The collected massive power grid operation data is processed and analyzed in real time by deploying distributed streaming computing platforms, such as APACHEKAFKA and APACHESPARKSTREAMING. And realizing real-time access of data by adopting ApacheKafkaConnect components, and converging the multi-source heterogeneous data streams into Kafka in a uniform format. Partitioning and persistent storage of data in Kafka ensures high throughput and fault tolerance of data.

And carrying out structural processing on the real-time data stream by APACHESPARKSQL, and realizing real-time conversion, filtering and aggregation of the data by defining a stream data table and continuously inquiring. By setting reasonable data windowing intervals (such as 5 seconds), quasi-real-time feature extraction and statistical index calculation such as average voltage, current peak value, frequency deviation and the like are realized, and data support is provided for subsequent abnormality detection and risk assessment. By adopting a high-reliability time synchronization mechanism, such as IEEE1588 Precision Time Protocol (PTP), a PTP master-slave clock is deployed at each node of a power system, GPS time service signals are utilized as synchronous information sources, so that the accurate synchronization of each node clock of the whole network is realized, the synchronization precision can reach a sub microsecond level, the consistency and comparability of collected data time stamps are ensured, and a foundation is laid for subsequent time sequence data analysis and causality excavation.

For the power grid operation data collected in real time, a multidimensional data quality monitoring and checking mechanism is constructed, a series of data quality indexes (such as sampling rate, time stamp continuity, numerical range and the like) are defined, the values of the indexes are calculated in real time by using a statistical method, and the values are compared with a preset threshold value, so that the problems of data loss, data abnormality and the like are found and processed in time. When the data is missing more than a certain proportion (such as 0.5%), automatically triggering a data repair flow, estimating and filling the missing data through algorithms such as interpolation, fitting and the like, and ensuring the integrity and accuracy of the data.

Based on the power grid operation data acquired and processed in real time, the historical data and expert knowledge are combined to construct a multi-model fusion fault diagnosis and risk early warning model. And automatically extracting key features in the data by adopting a time sequence feature extraction method, such as sliding window, wavelet transformation and the like, and constructing a high-dimensional feature vector. The machine learning algorithm, such as a support vector machine, a random forest and the like, is adopted to train a fault diagnosis model, and the model super-parameters are optimized through the methods of cross verification, grid search and the like, so that the fault diagnosis accuracy is improved. Aiming at risk early warning tasks, a deep learning model, such as a long and short time memory network (LSTM), is adopted to learn time sequence dependency relationship and abnormal modes in data, real-time prediction is carried out in a sliding window mode, and when the predicted risk probability exceeds a certain threshold value, alarm information is automatically generated. The prediction performance of the model is evaluated in real time, and the threshold value is dynamically adjusted by combining expert feedback, so that the early warning model is continuously optimized. The results of fault diagnosis and risk early warning are presented in real time through a visual instrument panel, and on-duty personnel are timely notified in a short message, mail and other modes, so that the safety and reliability of the operation of the power grid are improved.

For example, in order to realize high-performance data acquisition, a PCI-Express data acquisition card with 8 synchronous sampling channels can be selected, each channel is provided with a 24-bit ADC, the sampling rate is set to be 20kHz, acquired data is transmitted to a memory in real time in a DMA mode, and the data transmission delay can be controlled within 100 microseconds. And a voltage transformer and a current transformer are respectively arranged on the high-voltage side and the low-voltage side of the 220kV transformer substation, the transformation ratio of the transformers is 1000:1 and 500:1, and voltage and current signals are conditioned to be within +/-10V through a signal conditioning circuit and are matched with the measuring range of a data acquisition card.

The collected data is uploaded to a data center through industrial Ethernet, the Ethernet switch adopts gigabit optical fiber ports, the bandwidth can reach 1Gbps, and the transmission delay is less than 1 millisecond. The Kafka cluster is deployed in a data center and consists of 3 brooker nodes, each node is configured with a 32G memory and a 1TBSSD hard disk, and the nodes are interconnected through a 10Gbps network by adopting a rack-mounted server. The data retention time of Kafka was set to 7 days, the partition number was set to 16, and the collected data stream was imported into Kafka through KafkaConnect. The Spark cluster is composed of 1 master node and 4 worker nodes, each node is configured with 64G memory and 2TB hard disk, and HDFS is adopted as a distributed file system. The batch processing interval of SparkStreaming is set to be 5 seconds, a SparkSQL-defined streaming data table is used, real-time aggregation of data is realized through continuous inquiry, and statistical indexes such as average voltage, current and frequency are calculated every 5 minutes.

And a PTP master clock is deployed in the transformer substation, a time service signal is received through a GPS antenna, other intelligent equipment is used as a PTP slave clock, the PTP master clock is synchronized with the network, the synchronization period is set to be 1 second, and the clock deviation can be controlled within 1 microsecond. In the aspect of data quality monitoring, quality inspection is carried out on collected data every 1 minute, indexes such as mean value, variance, maximum value and minimum value of the data are calculated and compared with historical data, when the data quality is abnormal, abnormal data are automatically marked as invalid, a data restoration process is triggered, the invalid data are restored through Lagrange interpolation algorithm, and the size of an interpolation window is set to 10 data points.

And constructing a fault diagnosis model by using a support vector machine algorithm, wherein input characteristics comprise real-time indexes such as voltage, current, frequency and the like and environmental parameters such as temperature, humidity and the like, sample data are selected from historical data, and after data cleaning and characteristic engineering, model training and evaluation are performed by adopting a 10-fold cross validation mode, so that the accuracy of the model can reach more than 95%. For risk early warning, an LSTM model is adopted, data in the past 1 hour are used as input, risk probability in the future 1 hour is predicted, warning information is automatically generated when the risk probability exceeds 80%, and operators on duty are notified in a mail and short message mode, so that the running safety of a power grid is improved.

Further, a time synchronization technology is combined with power failure detection, the power grid voltage waveform is monitored in real time through a high-speed voltage sampling and edge detection algorithm, and the accurate moment of power failure is recorded.

In this embodiment, in order to realize high-precision time synchronization, the invention also adopts a high-precision GPS time synchronization technology, and by receiving GPS satellite signals, the system clock is synchronized in real time, so that the time synchronization precision is ensured to reach microsecond level, and a reliable time reference is provided for accurate time marking at the moment of power failure.

The invention combines the time synchronization technology with the power failure detection, monitors the voltage waveform of the power grid in real time through the high-speed voltage sampling and edge detection algorithm, immediately triggers the power failure event once the voltage dip or disappearance is detected, generates a time mark at microsecond response speed, and records the accurate moment of the occurrence of the power failure. The time mark of the power failure occurrence moment is recorded in a standardized time stamp format, including years, months, days, hours, minutes, seconds and milliseconds, so that the readability and the manageability of the time mark are ensured, and the subsequent data analysis and application are convenient. By adopting a redundant data storage mechanism, the power failure time mark is written into a plurality of independent storage media, such as a local Flash, a remote server database and the like, the data consistency among the plurality of storage media is ensured through a distributed consistency algorithm (such as Paxos or Raft), the safety and the reliability of data are improved, and the key data loss caused by single-point faults is prevented.

Aiming at recorded power outage time data, the time distribution, frequency, duration and the like of power outage events are counted and mined by applying big data analysis technologies such as time sequence analysis, an anomaly detection algorithm and the like, and potential rules and trends are found. And fusing the outage time data with other related data (such as weather conditions, load conditions and the like) to construct a multi-dimensional characteristic data set, and providing a rich data base for subsequent machine learning model training.

And a power outage prediction model is established by combining the historical power outage event data and related influence factors through a machine learning algorithm (such as an LSTM neural network), and the power outage risk in a certain future time range is estimated. The LSTM neural network can effectively capture long-term dependency relationship in time sequence data, and through learning of historical data, complex association between time characteristics and influence factors of power failure events is automatically extracted, so that prediction of future power failure risk is realized. By continuously training and optimizing the model, the accuracy and reliability of prediction are improved, decision basis is provided for preventive measures taken in advance by the power grid operation and maintenance department, the occurrence probability of power failure events is reduced, and the stability and reliability of the power grid are improved.

For example, in order to achieve high-precision time synchronization, a GPS time synchronization module, such as a NEO-M8N module, may be used to achieve synchronization with UTC (Coordinated Universal Time) time by receiving PPS (Pulse Per Second) signals broadcast by GPS satellites, where the time precision may be up to 100 nanoseconds. Meanwhile, the voltage of the power grid is sampled in real time through a voltage sampling module, the sampling frequency is set to be 10kHz, 100 voltage samples are collected in each period, frequency domain analysis is carried out on the voltage waveform through a Fast Fourier Transform (FFT) algorithm, and characteristic quantities such as fundamental wave amplitude and phase angle are extracted.

When the voltage dip exceeds a threshold (such as 50% of rated voltage) or the voltage is completely eliminated, the power failure event is judged to occur, and a power failure time tag is triggered to generate a power failure time tag, wherein the tag format is 'YYYY-MM-DD HH: MM: ss.SSS', and the power failure time tag is accurate to milliseconds. The power failure time tag is issued to the Kafka message queue through the MQTT protocol, then is synchronously written into a local SSD hard disk and a remote MySQL database through the Kafka Connect component, and data consistency among a plurality of storage nodes is realized through the Raft protocol, so that high reliability of data is ensured.

Aiming at stored massive power outage time data, a Facebook propset time sequence prediction algorithm is adopted, the time distribution of historical power outage events is modeled, the periodicity rule (such as daily, weekly, monthly and the like) of the occurrence of power outage is described, external factors such as weather conditions (such as temperature, humidity, wind speed and the like) and load conditions (such as current, power and the like) are introduced as regression variables, a multivariate time sequence prediction model is constructed, rolling prediction is carried out on the power outage risk in 7 days in the future, and early warning information is automatically generated when the predicted power outage probability exceeds 60%. By comparing and analyzing the model prediction result and the actual power failure situation, the model super-parameters (such as seasonal period, trend change rule and the like) are continuously optimized, the prediction accuracy is improved, and the accurate prediction and early warning of the power failure risk are realized.

In this embodiment, preferably, after collecting operation data of each node of the power system at the same time point, the method further includes:

In this embodiment, according to the synchronization time stamp obtained by the high-precision time synchronization technology, time calibration and alignment are performed on the power outage event data output by the data acquisition system, and by calculating the start-stop time difference of the power outage event, the accurate power outage duration is obtained, and aggregation statistics is performed according to different time scales of 15 minutes, 1 hour, 1 day and the like, so as to form a multi-granularity power outage duration distribution diagram. And (3) carrying out spatial correlation on the outage event data and a power grid topological structure by adopting a Geographic Information System (GIS) technology, rapidly positioning the influence range of the outage event through algorithms such as buffer zone analysis, spatial connection and the like, calculating key indexes such as the number of affected users, the power load and the like, and forming a visual thermodynamic diagram of the influence range of the outage.

Clustering and association analysis are carried out on the space-time distribution modes of the outage events through data mining and machine learning algorithms such as K-means clustering and Apriori association rule mining, rules of outage high-incidence areas, high-incidence periods and the like are found, external factors such as meteorological data and equipment operation data are combined, a outage risk prediction model is built, and quantitative evaluation and grading are carried out on outage risks in a future period of time. And dynamically adjusting the strategy and rule of the risk control system according to the power failure risk level, and realizing hierarchical control and self-adaptive response of the power failure risk by setting different thresholds and triggering conditions.

Aiming at the complexity of the risk control system, a Analytic Hierarchy Process (AHP) and a fuzzy comprehensive evaluation method are adopted to qualitatively and quantitatively evaluate factors such as the number of risk control strategies, the complexity of rules, the length of a decision chain and the like. The method comprises the steps of constructing a pair comparison matrix by establishing a three-level hierarchical structure comprising a target layer, a criterion layer and a scheme layer, calculating the maximum eigenvalue and consistency ratio of a judgment matrix, and determining the weight of each factor. And then, calculating the comment membership of each factor according to the membership function, constructing a fuzzy evaluation matrix, finally obtaining the comprehensive complexity index of the risk control system through weighted average synthesis, comparing with an industry reference value, and timely finding and adjusting an excessively complex or excessively simple risk control mechanism. Meanwhile, by adopting graph theory and complex network theory, modeling and analyzing the topological structure and association relation of the risk control system, evaluating the robustness and vulnerability of the risk control system by calculating the indexes such as average path length, clustering coefficient, centrality and the like of the network, optimizing and simplifying the structure, and improving the efficiency and reliability of risk control.

In the quantitative evaluation process, the method such as Monte Carlo simulation and sensitivity analysis is adopted to randomly sample and disturb key parameters such as time synchronization precision, data acquisition frequency, outage duration, outage influence range and the like for a plurality of times. And generating a large number of random sample points in the value range of the parameters through Latin hypercube sampling, and calculating a power failure processing time scale and a risk control complexity index corresponding to each sample point by using a simulation model to obtain a group of simulation results. And (3) evaluating the influence degree of each parameter on the simulation result by using a statistical method such as analysis of variance, correlation analysis and the like, calculating a Sobol sensitivity index, identifying the key parameter with the largest contribution to the model output, and providing decision support for parameter optimization and resource configuration. And feeding back the quantitatively evaluated fruiting time to a time synchronization system and a data acquisition system, and establishing a closed-loop feedback mechanism for power failure processing and risk control.

And dynamically adjusting parameters such as time synchronization precision, data acquisition frequency and the like according to the evaluation result, and continuously improving data quality and analysis precision. Meanwhile, the effect evaluation of risk control is fed back to a risk prediction model and a knowledge graph, and the risk control strategy and rules are continuously improved through incremental learning and parameter optimization to form a self-optimized and self-perfected risk management and control closed loop, so that the intelligent level and the actual effect of power failure treatment and risk control are continuously improved. By establishing a knowledge graph and a case library of the outage event, the reasons, influences, treatment processes and the like of the historical outage event are structurally represented and semantically associated, and a complete power outage coping knowledge system is formed. When a new power failure event occurs, a similar case is quickly searched through a knowledge reasoning and case matching algorithm, and a preliminary coping scheme and a decision proposal are automatically generated according to the handling experience and effect evaluation of the case, so as to assist operation and maintenance personnel in power failure handling and risk control.

In the space association analysis of the outage event data and the power grid topological structure, the outage event data and the power grid GIS data are firstly imported into an ArcGIS platform, a space connection tool is utilized, a outage event point is used as a target element, a power grid line and equipment are used as connection elements, two types of elements are spatially matched within a searching radius of 100 meters, and attribute information of the power grid elements, such as a line, equipment type, a power supply area and the like, is associated to the outage event point. And then, by analyzing the buffer areas, buffer areas with different radiuses such as 500 meters, 1 kilometer and 2 kilometers are generated by taking the power failure event point as the center, and the buffer areas are subjected to superposition analysis with a power user distribution diagram layer, so that the number and types of users in each buffer area are counted. Meanwhile, a spatial interpolation method, such as Kerling interpolation, is utilized to generate a spatial distribution grid layer of the power failure time according to the power failure time of the power failure event point, and the spatial distribution grid layer and the grid load distribution layer are subjected to weighted superposition to obtain the spatial distribution of the power load loss in the power failure period. And identifying the blackout high-incidence area and the blackout loss high-incidence area by using a hot spot analysis tool, for example Getis-OrdGi, generating a hot spot distribution map, and providing a space decision basis for blackout risk assessment and prevention and control.

In the uncertainty analysis of the quantitative evaluation key parameters, firstly, the distribution type and range of the parameters are estimated, for example, the time synchronization precision is subjected to normal distribution, the average value is 1 millisecond, the standard deviation is 0.2 millisecond, the data acquisition frequency is subjected to uniform distribution, the range is 10 to 60 seconds, the outage duration is subjected to exponential distribution, the average value is 30 minutes, the outage influence range is subjected to triangular distribution, the minimum value is 1 square kilometer, the most probable value is 5 square kilometers, and the maximum value is 10 square kilometers. Then, using Latin hypercube sampling, 1000 sample points are extracted within the distribution range of the parameters, generating an input matrix of the parameters. Then substituting the input matrix into a simulation model of the power failure processing time scale and the risk control complexity, and performing 1000 Monte Carlo simulations to obtain the distribution characteristics and the uncertainty range of the model output.

And (3) calculating a first-order index and a total index of each parameter for model output through sensitivity analysis, determining a key parameter with the greatest influence on output, and further analyzing a response relation curved surface between the key parameter and the output to provide a quantitative basis for parameter optimization configuration. The first-order index of the data acquisition frequency is 0.35, and the total index is 0.52, which shows that the influence on the power failure time scale statistics is the greatest, and the stability and the reliability of the power failure time scale statistics are ensured preferentially.

S2, based on a deep learning algorithm of a convolutional neural network, carrying out feature extraction and representation learning on the historical operation data, automatically learning a local mode and a global mode in the historical operation data through a convolutional layer and a pooling layer, and extracting depth feature vectors reflecting the intrinsic structure and time scale features of the data;

In this embodiment, a deep learning algorithm based on a Convolutional Neural Network (CNN) is adopted to perform feature extraction and representation learning on the acquired time sequence data, and a local mode and a global mode in the data are automatically learned through a convolutional layer and a pooling layer, so as to extract a deep feature vector reflecting the intrinsic structure and time scale features of the data.

The extracted depth feature vector is input into a multi-layer perceptron (MLP) network, the features are subjected to nonlinear transformation and classification through a full connection layer and an activation function, the MLP network is trained to identify and classify different types of abnormal modes, and the types and the confidence of the abnormal modes possibly existing in the data are obtained. And establishing a mapping relation between the abnormal mode and the risk level according to the domain knowledge and the historical data, and constructing a risk evaluation rule base.

Matching the identified abnormal pattern type and confidence with a risk assessment rule base, deducing the running state of the current system in real time, and dynamically assessing the potential risk level of the system according to the severity and duration of the abnormal pattern. And a rule-based reasoning engine is adopted to automatically generate risk early warning information and suggested countermeasures, and the running state and the risk level of the system are displayed in real time through a visual instrument panel, so that operation and maintenance personnel can know the health condition of the system in time.

Aiming at massive time sequence data and complex abnormal mode recognition tasks, a distributed computing framework, such as APACHESPARK and TensorFlow distributed version, is adopted to distribute data processing and model training tasks to a plurality of computing nodes for parallel execution, so that computing resources of clusters are fully utilized, and the efficiency of data processing and model training is improved. Through RDD (resilient distributed data set) and DAG (directed acyclic graph) scheduling mechanisms of Spark, reasonable data slicing and task scheduling strategies are designed, data transmission among nodes and load balancing of calculation tasks are optimized, and time delay and resource consumption of task execution are minimized. And adopting an increment learning strategy and an online learning strategy to continuously optimize and update the abnormal pattern recognition model.

The newly acquired data is input into the model in real time in a streaming mode, and the parameters of the model are finely adjusted by using a back propagation algorithm, so that the model can adapt to the dynamic change of the data, and the accuracy of abnormal pattern recognition is improved. Meanwhile, an active learning mode is adopted, sample data with the most representation and information quantity is selected through uncertainty sampling, high-quality training data is obtained through a manual labeling mode, offline training and optimization are regularly carried out on the model, and generalization performance of the model is improved.

Aiming at the problem of resource consumption of the algorithm, a model compression and acceleration technology is adopted, and the calculation cost and the storage cost of the algorithm are minimized on the premise of ensuring the abnormal pattern recognition performance. The complex deep learning model is converted into a lighter student model by a knowledge distillation technology, and the parameter quantity and the calculation complexity of the model are reduced by controlling the transmission of knowledge through a soft tag and temperature parameters. The model quantization technology is adopted, floating point operation in the model is converted into fixed point operation, and the numerical precision of the model is reduced and the calculation efficiency and the deployment speed of the model are improved through the quantization mapping of the weight and the activation value. In terms of algorithm performance evaluation and optimization, the abnormal pattern recognition algorithm is comprehensively evaluated from the dimensions of time complexity, space complexity, accuracy, recall rate and the like. The asymptotic time complexity of the large O sign analysis algorithm is adopted, and the calculation efficiency and the expandability of the algorithm are evaluated through the change of the data scale and the batch size. And a memory analysis tool is adopted to evaluate the memory occupation and the data read-write expense of an algorithm, and the memory layout and the access mode of the data are optimized. And through the modes of cross verification, leave-one-out method and the like, indexes such as the accuracy rate, recall rate, F1 value and the like of the algorithm on different data sets are evaluated, and the generalization capability and robustness of the algorithm are analyzed.

And aiming at performance bottlenecks and optimization points, measures such as feature selection, model cutting, parameter tuning and the like are adopted, so that the performance and the effect of the algorithm are continuously improved. And establishing a perfect model monitoring and evaluating mechanism, and continuously monitoring and evaluating the performance of the abnormal pattern recognition algorithm. Through reasonable design of an evaluation index system and a benchmark test scheme, the performance of the model is evaluated and compared regularly, problems existing in the model are found and positioned in time, and the performance and the effect of the abnormal pattern recognition algorithm are continuously improved through model optimization and algorithm improvement modes. Meanwhile, a model version management and rollback mechanism is established, iterative updating of the model is managed and tracked, and when a problem occurs in a new version model, the model can be rolled back to a stable version quickly, and continuous and stable operation of the power system is ensured.

For the implementation of the abnormal pattern recognition algorithm, the following method may be adopted, for example. A CNN model is used that contains 3 convolutional layers and 2 pooling layers, the first containing 16 3x3 convolutional kernels, the second containing 32 3x3 convolutional kernels, the third containing 64 3x3 convolutional kernels, with a ReLU activation function between the convolutional layers, the pooling layers using max pooling, step size 2. The output of the CNN model is characterized by a fully connected layer containing 128 neurons, and then classified by a Softmax layer containing N neurons, where N is the number of classes of abnormal patterns. Using the Adam optimizer, the learning rate was set to 0.001, the batch size was set to 128, and 100 epochs were trained.

In the construction process of the risk assessment rule base, 5 common abnormal modes including data mutation, data stagnation, data deletion, data abnormal fluctuation and data trend abnormality are summarized through analysis of abnormal modes in historical data, and the risk levels respectively correspond to 5 risk levels, and are sequentially 1-5 levels from low to high. The risk assessment rules are designed according to the confidence and duration of the abnormal pattern, such as when the confidence of the abnormal pattern exceeds 0.8 and the duration exceeds 10 minutes, the risk level is raised by 1 level.

In the design of a distributed computing framework, a Spark RDD programming model is adopted, time sequence data are partitioned according to time stamps, each partition contains 10 ten thousand pieces of data, data slicing is carried out through a Spark HashPartitioner, and the fact that data of the same equipment can be distributed into the same partition is guaranteed. In the implementation of the incremental learning and active learning strategy, every time 1 ten thousand new pieces of data are received, the model is incrementally trained once, the learning rate is set to 0.0001, the batch size is set to 256, and 10 epochs are trained. And adopting an uncertainty sampling strategy based on maximum entropy, selecting 1000 pieces of sample data with the lowest confidence coefficient, acquiring a label in a manual labeling mode, adding the label into a training set for offline training, performing offline training once a week, setting the learning rate to 0.0005, setting the batch size to 512, and training 50 epochs.

In the aspects of model compression and acceleration, a knowledge distillation technology is adopted, a lightweight CNN model comprising 2 convolution layers and 1 pooling layer is selected as a student model, the softmax layer temperature parameter of the teacher model is set to be 10, the temperature parameter of the student model is set to be 1, and 100 epochs are trained. And an 8-bit fixed-point quantization technology is adopted to quantize the weight and the activation value of the model, the volume of the quantized model is reduced by 4 times, and the reasoning speed is improved by 3 times.

During algorithm performance evaluation, a 10-fold cross validation mode is used for evaluating the accuracy rate, recall rate and F1 value of the model on an abnormal pattern recognition task, wherein the average accuracy rate reaches 95%, the average recall rate reaches 93%, and the average F1 value reaches 94%. The parameter quantity of the CNN model is reduced by 30% through the feature selection and model clipping technology, the reasoning time delay is reduced by 20%, and the instantaneity and the resource utilization efficiency of the algorithm are further improved.

In this embodiment, preferably, after extracting the depth feature vector reflecting the intrinsic structure and the time scale feature of the data, further includes:

In the embodiment, a time sequence anomaly detection algorithm based on deep learning is adopted to perform self-adaptive modeling and anomaly pattern recognition on system operation data acquired in real time, a time dependency relationship of time sequence data is modeled through a long-short-term memory network (LSTM), local features and global features of data are extracted by combining a Convolutional Neural Network (CNN), and accurate recognition on anomaly patterns is achieved.

In the model training process, a semi-supervised learning mode is adopted, a large amount of unlabeled historical data is used for pre-training, the normal mode of the data is initially learned through a loss function with minimized reconstruction errors, then fine adjustment is carried out through a small amount of marked abnormal data, the recognition capability of the model to the abnormal mode is improved through a cross entropy loss function and an Adam optimization algorithm, and the generalization performance of the model is evaluated through cross verification and early stop strategies.

Establishing a historical experience database based on a graph database, storing various abnormal modes and corresponding treatment schemes which occur in the running process of the system, and expressing the association relation and evolution rules among the abnormal modes through a graph structure to form an abnormal mode knowledge graph. When the algorithm identifies a new abnormal mode, searching similar or related abnormal modes in a historical experience database through a graph matching algorithm, calculating a similarity score between the new mode and the known mode by adopting measurement indexes such as cosine similarity or editing distance, if the highest score is lower than a preset threshold (e.g. 0.8), judging that the matching is failed, triggering a dynamic misjudgment risk assessment mechanism, and otherwise, directly calling a corresponding treatment scheme. Dynamic movement

The state misjudgment risk assessment mechanism carries out credibility assessment and risk analysis on the abnormal mode identified by the algorithm based on Bayesian reasoning and a decision tree algorithm. According to the characteristic attribute and the occurrence frequency of the abnormal mode, the posterior probability of the category to which the abnormal mode belongs is calculated by using a naive Bayesian algorithm to obtain a confidence score, and according to expert knowledge and experience data, reasonable prior probability and conditional probability distribution are set. And then, by combining the running state of the current system and environmental factors, evaluating the potential influence and the hazard degree of the abnormal mode on the system through a decision tree algorithm, generating a risk level by using a base index or an information gain ratio as a division criterion, and controlling the complexity and the overfitting risk of the tree through a pruning and optimizing algorithm. If the confidence score is below a threshold (e.g., 0.6) and the risk level is above a threshold (e.g., level 3), a false positive risk is determined, triggering an adaptive time scale adjustment mechanism. The adaptive time scale adjustment mechanism adapts to the time scale characteristics of different anomaly modes by dynamically adjusting the data sampling frequency and time window size.

When the misjudgment risk occurs, the time window is amplified to 2 times, the data sampling frequency is reduced to 1/2, more context information is acquired, and the long-term trend and the evolution rule of the abnormal mode are captured. If the misjudgment risk still exists after adjustment, the time window is further enlarged to 3 times, and the data sampling frequency is increased to 2 times, so that the fine granularity characteristic and the instantaneous change of the abnormal mode are obtained. If the misjudgment rate is reduced by not more than 5% after the time scale parameters are continuously adjusted for 3 times, triggering a manual intervention mechanism, automatically sending alarm information to operation and maintenance personnel, prompting suspicious abnormal modes and related data fragments, and requesting the operation and maintenance personnel to perform manual review and judgment. The operation and maintenance personnel can check the detailed information and the context data of the abnormal mode through the visualization tool, and judge whether the abnormal mode is true or not by combining own experience and domain knowledge. If the abnormal condition is confirmed, marking the abnormal condition as a new abnormal mode, adding the new abnormal mode into a historical experience database, and designing a corresponding treatment scheme; if the judgment is erroneous, the algorithm model and the rule threshold are further optimized, and the recognition accuracy is improved.

In the dynamic misjudgment risk assessment process, a closed-loop linkage mechanism of abnormal pattern recognition and risk assessment is established, the result time of abnormal pattern recognition is fed back to a risk assessment module, the prior probability and the conditional probability in the risk assessment model are dynamically updated, meanwhile, the result of risk assessment is fed back to the abnormal pattern recognition module, the similarity threshold value and the feature weight are dynamically adjusted, and real-time assessment and dynamic control of the system running risk are realized. In order to quantitatively evaluate the effectiveness of the dynamic misjudgment risk evaluation mechanism, a set of risk evaluation accuracy evaluation index system is designed, wherein the index system comprises key indexes such as abnormal pattern recognition rate, misjudgment rate, missed judgment rate, average response time and the like. And (3) carrying out off-line analysis and back measurement on system operation data at regular intervals, calculating the numerical values of various indexes by using assessment tools such as confusion matrix, ROC curve and the like, comparing the numerical values with a reference value, and timely finding and optimizing defects in an algorithm model and an assessment mechanism.

Illustratively, in the implementation of the abnormal pattern recognition algorithm, the following method may be employed. First, a historical dataset containing 1 ten thousand unlabeled samples and 1 thousand labeled samples was used, where the ratio of normal samples to abnormal samples was 9:1. In the pre-training phase, the LSTM self-encoder is used for reconstructing the label-free data, and the normal mode of the data is learned by minimizing reconstruction errors, so that 100 epochs are trained, and the learning rate is 0.001.

In the fine tuning stage, a CNN model is overlapped on the top of the LSTM, supervised learning is performed by using labeled data, abnormal recognition capability of the model is improved by minimizing cross entropy loss, 50 epochs are trained, and the learning rate is 0.0001. Through 5-fold cross validation, the generalization performance of the model is evaluated, and the average accuracy reaches more than 95%. In the construction of a historical experience database, a Neo4j graph database is used for modeling an abnormal mode and a treatment scheme as nodes and relations respectively, and graph traversal and matching are achieved through a Cypher query language.

When the abnormal pattern is matched, a cosine similarity algorithm is used for calculating a similarity score between the new pattern and the known pattern, and the threshold value is set to be 0.8. In the dynamic misjudgment risk assessment, a naive Bayesian algorithm is used for calculating posterior probability of an abnormal mode, prior probability is obtained according to historical data statistics, and each characteristic attribute of the conditional probability assumption is mutually independent. And evaluating the risk level of the abnormal mode by using a CART decision tree algorithm, dividing a feature space by using a base index, generating a decision tree with the maximum depth of 5, and optimizing by using a cost complexity pruning algorithm. The confidence score and risk level thresholds are set to 0.6 and 3 levels, respectively. In the adaptive time scale adjustment, the initial time window is 10 minutes and the data sampling frequency is 1 second. When the risk is misjudged, the time window is enlarged to 20 minutes, the sampling frequency is reduced to once 2 seconds, and if the risk still exists, the time window is enlarged to 30 minutes, and the sampling frequency is increased to once 0.5 seconds. After 3 continuous adjustments, if the false positive rate is reduced by no more than 5%, triggering a manual intervention. In a closed-loop linkage mechanism, an online learning algorithm such as AdaBoost and a random forest is used for updating a risk assessment model and an abnormal pattern recognition model in real time, and the self-adaption capability of the system is continuously improved through incremental training and parameter tuning.

In the evaluation index system, common indexes such as accuracy, precision, recall rate and F1 value are used, the common indexes are obtained through confusion matrix calculation, and the classification performance of the model is evaluated by using an ROC curve and an AUC value. Through regular offline analysis and manual feedback, the threshold setting is optimized, the abnormal pattern recognition rate is improved to be more than 99%, the misjudgment rate is controlled to be within 1%, the average response time is shortened to be within 10 seconds, and efficient and accurate dynamic misjudgment risk assessment is realized.

S3, establishing a power failure risk assessment model according to the depth feature vector, and optimizing the power failure risk assessment model by adopting a Bayesian network or Markov decision process method;

In this embodiment, preferably, the establishing a blackout risk assessment model according to the processed operation data, and optimizing the blackout risk assessment model by adopting a bayesian network or a markov decision process method includes:

In the embodiment, key parameters such as voltage, current, power and the like in the running process of the power grid are collected in real time through the high-speed data collection equipment and the intelligent sensor, time sequence data storage is carried out according to millisecond-level time granularity, and preprocessing operations such as cleaning, normalization and standardization are carried out on the data, so that the data quality and usability are improved. The method comprises the steps of carrying out feature extraction and dimension reduction compression on collected time sequence data by utilizing signal processing technologies such as wavelet transformation, fourier transformation and the like, extracting key indexes capable of reflecting power failure risk features such as voltage sag frequency, current distortion rate, harmonic content and the like, and constructing a multi-dimensional power failure risk assessment index system.

Aiming at blackout risk assessment indexes, adopting data mining and machine learning algorithms, such as association rule mining, time sequence anomaly detection and the like, analyzing space-time association relation and evolution rules among indexes, finding out precursor features and influence factors of blackout event occurrence, constructing a causal relation network model of blackout risk, and revealing an inherent mechanism of risk propagation. According to the causal relationship network of the outage risk, a Bayesian network reasoning algorithm is adopted to quantitatively calculate the prior probability and the conditional probability of each risk factor, probability parameters are dynamically updated through maximum likelihood estimation and an expected maximization algorithm, and real-time reasoning and prediction of the outage risk are realized.

Aiming at the time sequence characteristics and uncertainty of the power failure risk, a Markov decision process model is adopted to divide the power failure risk state into a plurality of levels such as normal, warning and serious, a transition probability matrix and an immediate compensation function between the risk states are defined, and an optimal risk decision strategy is solved through strategy iteration and value iteration algorithm, so that dynamic assessment and early warning of the power failure risk under different time scales are realized. According to the optimal strategy of the Markov decision process, risk control measures and emergency plans are designed, and a multi-scene and multi-level emergency response scheme is formulated.

The risk control strategy is continuously optimized and improved through reinforcement learning algorithms, such as Q learning and Sarsa algorithm, so that the effectiveness and reliability of risk prevention and control are improved. In order to adapt to dynamic change and uncertainty of a power grid operation environment, a self-adaptive dynamic Bayesian network, a non-stationary Markov decision process and other methods are adopted to learn and update a power failure risk assessment model on line. The self-adaptive dynamic Bayesian network dynamically adjusts the network structure and the conditional probability distribution by introducing time-varying parameters and a self-adaptive mechanism, and adapts to the change of data distribution. The non-stationary Markov decision process dynamically adjusts the optimal decision strategy by taking into account the state transition probabilities and the time-varying characteristics of the reward function, accommodating for environmental uncertainty. According to the newly acquired data and expert feedback, the model is optimized and updated in real time by utilizing technologies such as online gradient descent, incremental decision trees and the like, and the dynamic adaptability of the model is improved.

And constructing a closed-loop feedback mechanism of power failure risk assessment, and periodically evaluating and checking the effectiveness of the risk assessment model and method. And the sensitivity degree of the model to different parameters and factors is evaluated by adopting a local sensitivity analysis method and a global sensitivity analysis method, and the uncertainty and the reliability of the model are quantified by using technologies such as Monte Carlo simulation and the like. And introducing model stability indexes such as confidence intervals of model parameters, fluctuation rate of prediction errors and the like, and judging the stability and reliability of the model. And by combining analysis and summary of actual power failure events, the model is verified and calibrated, and the feedback result is utilized to drive parameter adjustment and structural optimization of the model, so that the risk function model is continuously improved and perfected, and the accuracy and the practicability of risk prediction are improved.

The method comprises the steps of establishing a visual simulation and early warning platform of the power failure risk, visually displaying distribution characteristics and evolution trend of the power failure risk in time and space dimensions through a data visual technology such as a thermodynamic diagram, a radar diagram and the like, setting a risk early warning threshold and a triggering rule, automatically generating risk early warning information and warning prompts when the risk level exceeds a preset threshold, and pushing the risk early warning information and the warning prompts to related responsible persons in real time through short messages, mails and the like, so that timeliness of risk perception and response is improved. And comparing and analyzing the risk assessment result with actual operation data, assessing pertinence and effectiveness of a risk control measure, and timely adjusting and optimizing the risk control measure to form a continuously improved and dynamically optimized risk management closed loop, so that reliable guarantee is provided for safe and stable operation of the power system. Through continuous iteration and perfection of the power failure risk function model, multi-time scale and omnibearing risk sensing and early warning of the power system are realized, and the safety, reliability and toughness of power grid operation are improved.

In the data acquisition and preprocessing stage, a high-speed data acquisition card, such as NIPXIe-6368, is used for sampling power grid operation parameters, the sampling rate is set to be 10kHz, the data bit width of each sampling channel is 16 bits, and real-time storage and management of mass time sequence data are realized by developing a data acquisition program through LabVIEW. And (3) carrying out wavelet denoising on the acquired data, selecting Daubechies4 wavelet basis function, enabling the decomposition layer number to be 5, carrying out soft threshold processing on the wavelet coefficient by a threshold method, and removing high-frequency noise. Then, the data is Z-score normalized, the data is mapped into the interval from 0 to 1, and the dimension effect is eliminated. And then, carrying out feature extraction on the data by adopting Principal Component Analysis (PCA), and selecting the first k principal components with the accumulated variance contribution rate of more than 95% as key features to realize dimension reduction compression of the data. In the causality mining stage, an Apriori association rule mining algorithm is adopted, the minimum support degree is set to be 0.05, the minimum confidence degree is set to be 0.8, association rules among power failure risk indexes are mined, and if 'voltage sag frequency is high and current distortion rate is greater than 5%, power failure risk grade is serious', and the like. Meanwhile, a Gaussian anomaly detection algorithm is adopted to perform anomaly detection on time series data, an anomaly threshold is set to be 3 times of standard deviation, and precursor features of occurrence of a power failure event are identified. In the Bayesian network construction stage, a TabuSearch algorithm is adopted to optimize a network structure, the length of a tabu table is set to be 10, the iteration number is 100, the network quality is estimated through a Minimum Description Length (MDL) score, an Expectation Maximization (EM) algorithm is adopted, a conditional probability table is iteratively calculated, and a convergence threshold is set to be 0.001.

In the optimization stage of the Markov decision process, a value iterative algorithm is adopted, a discount factor is set to be 0.9, a convergence threshold is set to be 0.0001, and an optimal state value function and a decision strategy are solved. In the self-adaptive dynamic adjustment stage, a Particle Swarm Optimization (PSO) algorithm is adopted, the particle number is set to be 50, the acceleration factor is 1.5, the inertia weight is 0.7, the parameters of the Bayesian network and the Markov decision process are optimized on line, and the dynamic change of data distribution is adapted.

In the sensitivity analysis stage, a Sobol global sensitivity analysis method is adopted to generate 5000 groups of Sobol sequences, first-order and total sensitivity indexes of each input parameter are calculated, importance and interaction of the parameters are evaluated, 1000 times of random sampling are carried out through Monte Carlo simulation, mean value, variance and quantile of model output are calculated, and uncertainty of a model is quantified.

In the aspect of visual early warning, a Grafana is adopted to build a real-time monitoring panel, the threshold value of the risk level is set to be 0.6, when the risk evaluation value exceeds the threshold value, an alarm is automatically triggered, and early warning information is pushed to the Web front end in real time through a WebSocket, so that real-time sensing and visualization of risks are realized.

And S4, acquiring current operation data of the power system, inputting the current operation data into an optimized power outage risk assessment model, outputting an assessment result, and constructing a power outage processing strategy matched with the assessment result and executed by a safety early warning terminal.

In this embodiment, preferably, the constructing a matched outage processing policy executed by the safety precaution terminal according to the output result of the optimized outage risk assessment model includes:

and constructing a power outage processing strategy executed by a safety precaution terminal based on the risk level, wherein the safety precaution terminal comprises at least one of the following:

In this embodiment, according to the power outage risk assessment result calculated by the risk function model, an equidistant dividing method is adopted to divide the risk value range into a plurality of levels such as low risk, medium risk, high risk, and extremely high risk. Wherein the risk value is between 0 and 1, 0 to 0.3 is low risk, 0.3 to 0.6 is medium risk, 0.6 to 0.8 is high risk, and 0.8 to 1 is extremely high risk. The dividing standard comprehensively considers factors such as safety margin of power grid operation, distribution characteristics of historical risk events and the like, and can reasonably reflect the level of risk. Weak links and key nodes in a power grid are identified through vulnerability analysis of a power system, indexes such as power flow distribution, static safety margin and the like of each node and each line are calculated, the influence degree of the power flow distribution, static safety margin and the like on the power failure risk is evaluated, power failure processing resources such as a standby power supply, an emergency power generation vehicle and the like are optimized and configured accordingly, and the resistance risk capacity of the power grid is improved. And (3) monitoring an output result of the risk function model in real time, dynamically tracking the change trend of the risk value, and automatically triggering an early warning mechanism when the risk value exceeds a preset threshold. Different early warning strategies and notification modes are adopted according to the level of the risk level.

And triggering a telephone notification and lifting the emergency plan to the highest priority on the basis of the audible and visual alarm and the short message notification for the high risk level. According to the risk grade and the influence range, corresponding operation and maintenance team and professional are automatically matched, detailed risk analysis reports and treatment suggestions are pushed, and the operation and maintenance personnel are guided to develop risk prevention and control and emergency treatment work. And aiming at risk assessment results of different time scales, flexibly adjusting the time window and granularity of power failure processing, and realizing risk analysis of multiple time scales. If the short-term risk value within 1 hour exceeds the threshold value, the original data are aggregated into hour granularity data according to minutes, the sensitivity and instantaneity of risk perception are improved, and if the long-term risk value over 1 day exceeds the threshold value, the hour granularity data are aggregated into day granularity data, and the long-term trend and the cumulative effect of risks are captured. And the depth prevention and control and panoramic perception of risks are realized through dynamic adjustment of time scales.

And automatically generating a multi-scene and multi-level emergency plan and a treatment plan according to the risk assessment result and the vulnerability analysis of the power system. By adopting case reasoning and knowledge graph technology, similar historical risk scenes and handling experiences are matched, a targeted emergency plan is dynamically generated by combining the current power grid running state and load characteristics, the emergency plan comprises load cutting, equipment isolation, scheduling strategies and the like, and feasibility and effectiveness of the emergency plan are optimized through simulation deduction and evaluation. In the emergency plan execution process, the running state and risk level change of the power grid are monitored in real time, the effect and influence of emergency measures are evaluated through real-time analysis of fault information, alarm events, operation logs and other data, and the emergency strategies are dynamically optimized and adjusted according to feedback results, such as load removal proportion adjustment, scheduling scheme modification and the like, so that the accuracy and adaptability of emergency treatment are ensured. And (3) performing closed loop control on risk assessment and emergency treatment, and integrating a risk function model with an emergency command scheduling system to realize full-flow automation of risk perception, assessment, early warning, decision making, execution and feedback.

The risk threshold value is adjusted in a self-adaptive mode through a reinforcement learning algorithm, super-parameters of a risk function model are optimized through a Bayesian optimization algorithm, an emergency knowledge base is updated dynamically through an incremental learning algorithm, and the intelligent level and instantaneity of risk management and control are improved continuously.

In the aspect of dynamic adjustment of the risk threshold, an adaptive mechanism is introduced, and the risk threshold is dynamically adjusted according to historical data and expert experience so as to cope with the change of the power grid operation condition and the iterative update of the risk assessment model. The dynamic threshold value of the risk value is calculated in real time by adopting methods such as sliding window, weighted average and the like, and the self-adaptive optimization of the risk judgment standard is realized through the dynamic adjustment of the threshold value, so that the flexibility and the robustness of the threshold value setting are improved. The automatic response flow after the risk threshold is exceeded is further refined, and the automatic response flow comprises generation, transmission, upgrading and releasing of early warning information, task allocation, scheduling and coordination of operation and maintenance personnel, starting, executing and tracking of emergency measures and the like. The workflow engine and the task scheduling system realize the automatic arrangement and execution of the response flow, and adopt a real-time monitoring and exception handling mechanism to ensure the reliability and continuity of the response process. Meanwhile, a plan knowledge base and an expert decision support system are established, real-time assistance and knowledge enabling are provided for operation and maintenance personnel, and man-machine cooperation efficiency and emergency treatment capacity are improved.

The risk management and control visual platform for operation and maintenance personnel is constructed, information such as risk distribution, emergency resource allocation, treatment task execution and the like of a power grid is intuitively displayed through technologies such as an instrument panel and a geographic information system, interactive decision support and expert auxiliary functions are provided, man-machine coordination is enhanced, and efficiency and accuracy of risk management and control are improved. Meanwhile, experience data and case knowledge of risk management and control are accumulated to form a reusable and movable risk management knowledge base, and data support and knowledge accumulation are provided for subsequent model optimization and decision improvement.

The risk value output by the risk function model is normalized in the risk classification process, a Min-Max standardization method is adopted to map the risk value into a 0-1 interval, then the 0-1 interval is equally divided into 4 subintervals according to an equidistant classification principle, the low risk interval is [0,0.3], the medium risk interval is (0.3, 0.6], the high risk interval is (0.6,0.8), and the extremely high risk interval is (0.8,1), in the aspects of real-time monitoring and early warning, a sliding window algorithm is adopted, the window size is set to be 1 hour, the step length is 1 minute, the average risk value in each window is calculated in real time, and when the average risk value exceeds a preset threshold value (such as 0.7), an early warning mechanism is triggered.

For the extremely high risk level, the early warning notification is sent once every 5 minutes in a short message and mail mode, and on the basis of the short message and mail, the early warning notification is sent once every 1 minute in a telephone voice broadcasting mode, and meanwhile, the early warning information is pushed to a mobile terminal APP of an operation and maintenance person, and timeliness and arrival rate of the early warning information are guaranteed in the modes of message pushing, popup window reminding and the like of the APP. In the aspect of emergency plan generation and optimization, a multi-objective optimization model based on a genetic algorithm is adopted, power failure range minimization and power supply recovery time minimization are taken as optimization targets, load cut-off proportion, equipment isolation state, scheduling strategy and the like are taken as decision variables, and an optimal emergency decision combination is searched through operations such as selection, intersection, variation and the like of the genetic algorithm. In each iteration, calculating the fitness value of the emergency decision combination through network breaking simulation and risk assessment, updating the genetic codes of individuals in the population according to the fitness value, and continuously optimizing the search direction until a global optimal solution meeting constraint conditions is found out, so as to form a final emergency plan.

In knowledge base construction and learning, ontology-based knowledge representation and reasoning technology is adopted to construct a risk management and control domain ontology, knowledge such as a risk scene, an assessment model, an emergency plan and the like is expressed in a formal mode and semantically organized, ontology knowledge is stored by adopting a graph-based database (such as Neo4 j), and knowledge retrieval and reasoning are achieved through SPARQL query language.

In the expert system, an reasoning engine based on rules is adopted to encode expert experience and judgment rules to form a decision rule base for risk management and control, and the functions of risk assessment, early warning decision, emergency assistance and the like are realized through forward reasoning and reverse reasoning. Meanwhile, machine learning technologies such as incremental learning and active learning are adopted, and a knowledge base and a rule base are continuously optimized and expanded through expert feedback and case accumulation, so that the self-learning and evolution capability of the system is improved.

As shown in fig. 2, another embodiment of the present invention provides a power outage processing apparatus in a power protection area, which includes an acquisition module 11, an extraction module 12, an evaluation module 13, and an output module 14, wherein,

The acquisition module 11 is used for carrying out time marking at the moment of power failure according to a high-precision time synchronization method and acquiring historical operation data of each node of the power system at the same time point;

the extracting module 12 is used for extracting features and performing representation learning on the historical operation data based on a deep learning algorithm of the convolutional neural network, automatically learning a local mode and a global mode in the historical operation data through a convolutional layer and a pooling layer, and extracting depth feature vectors reflecting the intrinsic structure and time scale features of the data;

The evaluation module 13 is configured to establish a blackout risk evaluation model according to the depth feature vector, and optimize the blackout risk evaluation model by adopting a bayesian network or a markov decision process method;

And the output module 14 is used for acquiring the current operation data of the power system, inputting the current operation data into the optimized power outage risk assessment model, outputting an assessment result, and constructing a power outage processing strategy matched with the assessment result and executed by the safety early warning terminal.

As shown in fig. 3, another embodiment of the present invention provides a power outage processing apparatus for a power outage area, including a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21, where the processor 21 implements steps in an embodiment of a power outage processing method for the power outage area, such as steps S1 to S4 in fig. 1, when executing the computer program, or where the processor 21 implements functions of modules in the embodiments of the apparatus, such as the acquisition module 11, the extraction module 12, the evaluation module 13, and the output module 14, when executing the computer program.

Illustratively, the computer program may be split into one or more modules that are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a particular function for describing the execution of the computer program in a power outage handling apparatus in the power assurance area.

The power outage handling equipment in the power protection area may include, but is not limited to, a processor 21, a memory 22. It will be appreciated by those skilled in the art that the schematic diagram is merely an example of a power outage handling apparatus in a power outage area, and does not constitute a limitation of the power outage handling apparatus in the power outage area, and may include more or less components than those illustrated, or may be combined with certain components, or different components, e.g., the power outage handling apparatus in the power outage area may further include input and output devices, network access devices, buses, etc.

The Processor 21 may be a central processing unit (Central Processing Unit, CPU), but may also be other general purpose processors, digital signal processors (DIGITAL SIGNAL Processor, DSP), application SPECIFIC INTEGRATED Circuit (ASIC), off-the-shelf Programmable gate array (Field-Programmable GATE ARRAY, FPGA) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc., and the processor 21 is a control center for power outage processing in the power protection area, and connects various parts of power outage processing equipment in the entire power protection area by using various interfaces and lines.

The memory 22 may be used to store the computer program and/or module, and the processor 21 may implement various functions of the power outage handling of the power protection area by running or executing the computer program and/or module stored in the memory 22 and invoking data stored in the memory 22. The memory 22 may mainly include a storage program area that may store an operating system, application programs required for at least one function (such as a sound playing function, an image playing function, etc.), etc., and a storage data area that may store data created according to the use of the cellular phone (such as audio data, a phonebook, etc.), etc. In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, smart memory card (SMART MEDIA CARD, SMC), secure Digital (SD) card, flash memory card (FLASH CARD), at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device.

Wherein the integrated module for power outage handling in the power assurance area may be stored in a computer readable storage medium if implemented as a software functional unit and sold or used as a stand alone product. Based on such understanding, the present invention may implement all or part of the flow of the method of the above embodiment, or may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of each of the method embodiments described above. Wherein the computer program comprises computer program code which may be in source code form, object code form, executable file or some intermediate form etc. The computer readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth.

Those skilled in the art will appreciate that implementing all or part of the above-described methods in accordance with the embodiments may be accomplished by way of a computer program stored on a computer readable storage medium, which when executed may comprise the steps of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a random-access Memory (Random Access Memory, RAM), or the like.

Accordingly, an embodiment of the present invention provides a computer readable storage medium, where the computer readable storage medium includes a stored computer program, and when the computer program runs, the device where the computer readable storage medium is located is controlled to execute steps in the power outage processing method in the power protection area according to the foregoing embodiment, for example, steps S1 to S4 described in fig. 1.

The embodiment of the invention has the beneficial effects that at least one point of the following is:

The foregoing examples illustrate only a few embodiments of the invention and are described in detail herein without thereby limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention. Accordingly, the scope of protection of the present invention is to be determined by the appended claims.

Claims

1. The power failure processing method for the power guarantee area is characterized by comprising the following steps of:

The method comprises the steps of combining a time synchronization technology with power failure detection, monitoring a power grid voltage waveform in real time through a high-speed voltage sampling and edge detection algorithm, and recording the accurate moment of power failure occurrence;

The method comprises the steps of carrying out feature extraction and representation learning on historical operation data based on a deep learning algorithm of a convolutional neural network, automatically learning a local mode and a global mode in the historical operation data through a convolutional layer and a pooling layer, extracting depth feature vectors reflecting intrinsic structure and time scale features of the data, and after extracting the depth feature vectors reflecting the intrinsic structure and time scale features of the data, further comprising:

Establishing a historical experience database based on a graph database, and acquiring various abnormal mode types and corresponding treatment schemes which occur in the running process of a storage system, wherein the abnormal mode types comprise data mutation, data stagnation, data deletion, data abnormal fluctuation and data trend abnormality;

Acquiring current operation data, inputting the current operation data into the anomaly identification model, carrying out matching processing on the new anomaly mode based on a graph matching algorithm after identifying the new anomaly mode, triggering a dynamic erroneous judgment risk assessment mechanism if matching fails, otherwise, directly calling a corresponding treatment scheme, and carrying out matching processing on the new anomaly mode based on the graph matching algorithm, wherein the method comprises the following steps:

Searching similar or related abnormal modes in the historical experience database through a graph matching algorithm, calculating similarity scores of the new abnormal modes and each abnormal mode type, if the highest score is lower than a first preset threshold, matching fails, triggering a dynamic misjudgment risk assessment mechanism, otherwise, directly calling a corresponding treatment scheme, wherein the dynamic misjudgment risk assessment mechanism comprises calculating confidence and risk levels of the new abnormal modes, if the confidence is lower than a second preset threshold and the risk levels are higher than the preset threshold, misjudgment risks and triggering an adaptive time scale adjustment mechanism, and the adaptive time scale adjustment mechanism comprises judging whether the new abnormal modes are truly abnormal or not through dynamically adjusting data sampling frequency and time window size, and if the new abnormal modes are truly abnormal, designing a treatment scheme corresponding to the new abnormal modes;

establishing a blackout risk assessment model, and optimizing the blackout risk assessment model by adopting a Bayesian network or a Markov decision process method, wherein the establishing the blackout risk assessment model, and optimizing the blackout risk assessment model by adopting the Bayesian network or the Markov decision process method comprises the following steps:

According to the blackout risk assessment model, a Bayesian network reasoning algorithm is adopted to quantitatively calculate probability parameters of various risk factors output by the blackout risk assessment model, and the probability parameters are dynamically updated through maximum likelihood estimation and an expected maximization algorithm so as to realize real-time reasoning and prediction of blackout risk;

the method comprises the steps of obtaining current operation data of the power system, inputting the current operation data into an optimized power failure risk assessment model, outputting an assessment result, constructing a power failure processing strategy matched with the assessment result and executed by a safety early warning terminal, and flexibly adjusting a time window and granularity of power failure processing according to the assessment results of different time scales after outputting the assessment result so as to realize risk analysis of multiple time scales.

2. The power outage handling method for a power management area according to claim 1, further comprising, after collecting historical operation data for each node of the power system at the same point in time:

3. The power outage handling method for the power protection area according to claim 1, wherein the constructing the power outage handling policy performed by the safety precaution terminal and matched with the evaluation result comprises:

4. The power outage handling method of the power outage guarantee area of claim 3, wherein the safety precaution terminal comprises at least one of:

5. A power outage handling apparatus for a power protection area, comprising:

the system comprises a power failure detection module, a power grid voltage waveform acquisition module and a power grid voltage waveform acquisition module, wherein the power grid voltage waveform acquisition module is used for carrying out time marking at the moment when a power failure occurs according to a high-precision time synchronization method and acquiring historical operation data of all nodes of a power system at the same time point;

The extraction module is used for carrying out feature extraction and representation learning on the historical operation data based on a deep learning algorithm of the convolutional neural network, automatically learning a local mode and a global mode in the historical operation data through a convolutional layer and a pooling layer, and extracting depth feature vectors reflecting the intrinsic structure and time scale features of the data; after extracting depth feature vectors reflecting the data-internal structure and time-scale features, further comprising:

the power failure risk assessment module is used for establishing a power failure risk assessment model and optimizing the power failure risk assessment model by adopting a Bayesian network or Markov decision process method, wherein the power failure risk assessment model establishment comprises the following steps of:

The power failure risk analysis system comprises an output module, an evaluation result and a power failure processing strategy, wherein the output module is used for acquiring current operation data of the power system, inputting the current operation data into an optimized power failure risk evaluation model, outputting the evaluation result, constructing a power failure processing strategy matched with the evaluation result and executed by a safety early warning terminal, and flexibly adjusting a time window and granularity of power failure processing according to the evaluation results of different time scales after outputting the evaluation result so as to realize risk analysis of multiple time scales.

6. A power outage handling apparatus for a power outage area, comprising a memory, a processor and a computer program stored on said memory and executable on said processor, said processor implementing the power outage handling method for a power outage area according to any one of claims 1 to 4 when executing said computer program.

7. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the power outage handling method for a power protection area according to any one of claims 1 to 4.