Search

CN-121980481-A - Power system data anomaly identification and active protection method and system based on deep learning

CN121980481ACN 121980481 ACN121980481 ACN 121980481ACN-121980481-A

Abstract

The invention provides a deep learning-based power system data anomaly identification and active protection method and system, which relate to the technical field of power system safety protection and comprise the following steps of collecting measurement data, switch state data and operation environment data of each node of a power system, and preprocessing the data; the invention adopts a multi-mode generation countermeasure network architecture of the fusion CNN and the LSTM to capture the space-time joint distribution characteristics of the power data, combines the Wasserstein distance to realize the distribution contrast of the real data and the reconstruction data, can effectively detect the refined false data injection attack which is difficult to find by the traditional method and keeps the statistical mean unchanged, improves the anomaly detection accuracy, reduces the false omission rate and solves the problem of insufficient anomaly identification precision in the prior art.

Inventors

  • HE LIPING
  • Zhu Fengzeng

Assignees

  • 临沂大学

Dates

Publication Date
20260505
Application Date
20260408

Claims (10)

  1. 1. The power system data anomaly identification and active protection method based on deep learning is characterized by comprising the following steps of: s1, collecting measurement data, switch state data and running environment data of each node of a power system, and preprocessing the data; s2, constructing a space-time feature consistency check mechanism based on a generating countermeasure network GAN architecture of a fused convolutional neural network CNN and a long-short term memory network LSTM, capturing space-time joint distribution features of data, calculating Wasserstein distances of real data distribution and reconstructed data distribution, and identifying abnormal data which do not accord with physical space-time relevance; S3, constructing a dynamic graph neural network DYNAMIC GNN model, updating an adjacent matrix in real time according to the switch state data, learning the dynamic association weights among nodes by using a graph attention mechanism, and positioning an abnormal starting node and a propagation path by combining the contribution degree of a shape value quantization node to abnormal loss; S4, integrating SHAP value analysis frames, quantifying contribution degree of each input feature to anomaly judgment, constructing an anomaly propagation path tracking algorithm, generating a visual anomaly report, and supporting matching and searching of historical anomaly cases; s5, dividing the abnormality into three classes according to the tracing result of the abnormality root cause, and generating and executing a hierarchical self-adaptive active protection strategy by combining the partition protection requirement of the power monitoring system; and S6, adopting an edge-cloud cooperative framework to realize cooperative execution of the edge node lightweight task and the cloud heavy task and continuous optimization of the model.
  2. 2. The method for identifying and actively protecting data anomalies of a power system based on deep learning as set forth in claim 1, wherein in S1, the data preprocessing comprises data cleaning, normalization and missing value filling, wherein the missing value filling adopts an interpolation algorithm based on LSTM, and the specific formula is: , Wherein, the Filling the result for the missing value at the time t of the ith node, As the weight of the kth history data, And (3) the historical measurement data of the ith node t-k moment, n is the historical data window length, and b is the bias term.
  3. 3. The method for identifying and actively protecting data anomalies of a power system based on deep learning as set forth in claim 1, wherein in S2, the calculation formula of the Wasserstein distance is: , wherein W (P, Q) is the Wasserstein distance between the true data distribution P and the reconstructed data distribution Q, For all sets of joint distributions from the true data distribution P to the reconstructed data distribution Q, Is a sample space of the real data and, The sample space of the reconstructed data is reconstructed, In order to be a sample of the real data, The data samples are reconstructed and the data samples are then processed, For the sample And (3) with Is used for the distance of euclidean distance, For joint distribution At the position of Probability density at.
  4. 4. The method for identifying and actively protecting data anomalies of a power system based on deep learning as set forth in claim 1, wherein in S3, the calculation formula of the shape value is: , Wherein, the The method comprises the steps of determining a Shapley value of an ith node, namely the contribution degree of the node i to abnormal loss, wherein N is a set of all nodes of the power system, S is any node subset which does not contain the node i, and N is the total number of nodes; for the abnormal loss function value of the node subset S, Is an abnormal loss function value of the node subset S { i } including the node i.
  5. 5. The method for identifying and actively protecting data anomalies in a deep learning-based power system according to claim 1, wherein in S5, the decision formula of anomaly classification is: , The method comprises the steps of determining an abnormality level, wherein L is an abnormality level determination value, alpha and beta are abnormality type weights and abnormality degree weights respectively, alpha+beta= 1;T is an abnormality type coefficient, equipment faults T=1, data interference T=2, external attack T=3, D is an abnormality degree coefficient, general level D=1, serious level D=2 and urgent level D=3, and determining the abnormality level according to the value range of L, wherein L is 1-2 and is 2-3, L is 2-3 is a serious level, and L is 3-5 is an urgent level.
  6. 6. The method for identifying and actively protecting data anomalies of a power system based on deep learning as set forth in claim 1, wherein in S6, the data interaction delay between the edge node and the cloud node satisfies the following formula: , Wherein, the In order to achieve a total interaction delay, The data transfer delay for the 5g+mec technique, Delay for node data processing, and 。
  7. 7. The power system data abnormality identification and active protection system based on deep learning is applied to the power system data abnormality identification and active protection method based on deep learning as set forth in any one of the claims 1-6, and is characterized by comprising a data acquisition module, a space-time feature verification module, a topology self-adaptive positioning module, an interpretability tracing module, a hierarchical protection module and an edge-cloud cooperation module; The system comprises a data acquisition module, a time-space feature verification module, a topology self-adaptive positioning module, an interpretive tracing module, a grading protection module, a data interaction and model collaborative optimization module, wherein the data acquisition module is used for acquiring measurement data, switch state data and running environment data of each node of a power system and preprocessing, the time-space feature verification module is used for constructing a time-space feature consistency verification mechanism based on a GAN framework fusing CNN and LSTM, calculating Wasserstein distance identification abnormal data, the topology self-adaptive positioning module is used for constructing DYNAMIC GNN models, updating an adjacent matrix in real time and combining with Shapley values to position abnormal initial nodes and propagation paths, the interpretive tracing module is used for integrating SHAP value analysis frameworks to generate visual abnormal reports and support historical case retrieval, the grading protection module is used for combining partition protection requirements according to abnormal classification grading results, and the edge-cloud cooperation module is used for deploying edge nodes and cloud nodes, and achieving data interaction and model collaborative optimization through the 5G+MEC technology.
  8. 8. The deep learning-based power system data anomaly identification and active protection system of claim 7, wherein the space-time feature verification module comprises a generator and a discriminator, the generator is composed of a CNN layer and an LSTM layer in series, the CNN layer is used for extracting spatial features of data, the LSTM layer is used for extracting temporal features of data, the discriminator adopts a fully connected neural network and is used for distinguishing real data from reconstruction data output by the generator, and a loss function of the discriminator is related to Wasserstein distance.
  9. 9. The deep learning-based power system data anomaly identification and active protection system according to claim 7, wherein the hierarchical protection module comprises an anomaly classification unit, a level judgment unit and a strategy generation unit, the anomaly classification unit classifies anomalies into equipment failure types, data interference types and external attack types according to anomaly root causes, the level judgment unit is used for calculating anomaly levels, and the strategy generation unit is used for generating protection strategies from a data level, an equipment level and a network level respectively aiming at anomalies of different types and levels and adopting differential protection measures in combination with safety level requirements of a production control area and a management information area.
  10. 10. The deep learning-based power system data anomaly identification and active protection system of claim 7, wherein the edge-cloud cooperation module comprises edge nodes and cloud nodes, the edge nodes are used for deploying data preprocessing, real-time anomaly identification and local protection to execute light tasks, the cloud nodes are used for deploying model training, global root cause analysis, strategy optimization and data storage heavy tasks, and the cloud nodes are used for iteratively optimizing anomaly identification models and protection strategies by receiving real-time data uploaded by the edge nodes and synchronizing the optimized models and strategies to all the edge nodes to realize cooperative linkage of edges and the cloud.

Description

Power system data anomaly identification and active protection method and system based on deep learning Technical Field The invention relates to the technical field of safety protection of power systems, in particular to a method and a system for identifying and actively protecting data anomalies of a power system based on deep learning. Background Along with the rapid development of the intelligent power grid, the power system is gradually transformed into digital, intelligent and networked power systems, and a large number of measuring equipment, monitoring terminals and communication modules are deployed, so that the data volume of the power system is increased in an explosive manner, and the accuracy and the integrity of the data directly determine the scientificity and the safety of the scheduling decision of the power system. At present, the data anomaly identification of the power system mainly adopts a traditional threshold detection method, a statistical analysis method and a simple machine learning method, and is matched with a passive protection strategy and a centralized system architecture, wherein the anomaly identification method mainly identifies obvious anomalies in data by setting a fixed threshold or statistical characteristics, the protection strategy is mainly unified passive response, and the system architecture mainly adopts cloud centralized processing and is responsible for data storage, model calculation and protection instruction issuing. In addition, some prior art attempts to introduce a single deep learning model for data anomaly detection, and combine a simple graph model to realize node positioning, so that a certain technical support is provided for the data security of the power system; However, in the prior art, the anomaly identification precision is insufficient, the traditional method and a single deep learning model cannot capture the space-time joint distribution characteristics of the electric power data, the refined false data injection attack with unchanged statistical mean value is difficult to detect, the adaptability to the change of a topological structure is poor, false alarm is easy to occur, the anomaly positioning is fuzzy, the traditional positioning method cannot adaptively track the dynamic change of the power grid topology, the anomaly starting node and the propagation path are difficult to accurately position, great inconvenience is brought to the investigation of operation and maintenance personnel, the interpretability is lacking, the traditional anomaly identification model is mostly a 'black box' model, the contribution degree of each characteristic to anomaly judgment cannot be quantized, the operation and maintenance personnel cannot understand the cause of anomaly generation, the decision efficiency is influenced, the protection strategy is stiff, the traditional passive protection mode cannot realize accurate matching according to the anomaly type, degree and cause, the protection response is delayed, the anomaly expansion is difficult to be restrained, the system real-time performance and reliability are insufficient, the centralized architecture causes the abnormal response of the edge node to be delayed, and the data transmission delay is high, and the real-time protection requirement of a large-scale intelligent power grid cannot be met. These drawbacks seriously affect the safe and stable operation of the power system, and there is a need for a highly efficient, accurate, interpretable and adaptive abnormality identification and active protection technique and system, so that the invention proposes a deep learning-based power system data abnormality identification and active protection method and system to solve the problems in the prior art. Disclosure of Invention Aiming at the problems, the invention provides a deep learning-based power system data anomaly identification and active protection method and system, adopts a multimode generation countermeasure network architecture fused with CNN and LSTM, captures the space-time joint distribution characteristics of power data, combines the Wasserstein distance to realize the distribution contrast of real data and reconstruction data, can effectively detect the refined false data injection attack which is difficult to find by the traditional method and keeps the statistical mean unchanged, improves the anomaly detection accuracy, reduces the report missing rate, and solves the problem of insufficient anomaly identification accuracy in the prior art. The invention aims at realizing the technical scheme that the method for identifying and actively protecting the data anomalies of the electric power system based on deep learning comprises the following steps: s1, collecting measurement data, switch state data and running environment data of each node of a power system, and preprocessing the data; s2, constructing a space-time feature consistency check mechanism based on a generating countermeas