Power grid fault disposal method and system based on multi-agent cooperation and knowledge graph driving

Abstract

A method and system for power grid fault handling based on multi-agent collaboration and knowledge graph-driven approach is disclosed. The method includes fusing multi-source power grid fault data, performing real-time monitoring and fault risk analysis of the power grid operating status based on the fusion results, including analysis of cross-section overload, equipment overload, weak points in the power grid, and power outage analysis of main transformers or busbars; automatically identifying and extracting entity knowledge from historical contingency plan text data for power grid fault handling based on a BiLSTM-CRF deep learning model and a TextCNN text convolutional neural network, and constructing a power grid fault handling knowledge graph; using a RE2 text similarity matching model to perform intelligent text matching of the fault risk analysis and the power grid fault handling knowledge graph; and employing a multi-agent reinforcement learning algorithm based on the Actor-Critic framework to generate the optimal handling strategy for power grid faults. This invention improves decision-making efficiency and system collaboration capabilities.

Description

Technical Field

[0001] This invention belongs to the field of power grid fault handling, and specifically relates to a power grid fault handling method and system based on multi-agent collaboration and knowledge graph-driven approach. Background Technology

[0002] With the deepening of the construction of new power systems, the power grid structure is becoming increasingly complex and its operation modes are becoming more diversified. New characteristics such as the high proportion of renewable energy grid integration, the large-scale application of power electronic equipment, and multi-dimensional interaction between power sources, grids, loads, and storage have led to new features in power grid faults, including rapid propagation, wide impact range, and complex coupling relationships. At the same time, the requirements for power supply reliability are constantly increasing. The latest plans require fault handling time in critical load areas to be controlled within milliseconds, posing a severe challenge to traditional fault handling technologies. Currently, power grid fault handling mainly relies on dispatchers manually analyzing fragmented multi-source information and making decisions based on personal experience. This results in problems such as difficulty in information fusion, low handling efficiency, weak coordination capabilities, and insufficient knowledge transfer, making it difficult to meet the requirements of modern power grids for rapid, accurate, and reliable intelligent fault handling.

[0003] The current field of power grid fault handling faces three major technical challenges: First, centralized decision-making architectures struggle to adapt to the rapid development of distributed energy resources, and the master station system faces computational bottlenecks when processing massive amounts of heterogeneous data. Existing systems struggle to effectively integrate multi-source heterogeneous data from SCADA, WAMS, protection devices, and environmental monitoring, resulting in fragmented fault information and inconsistent representations, hindering real-time and accurate perception of the power grid's operational status and dynamic risk quantification assessment. Second, static knowledge representation methods cannot accurately depict the complex relationships in the dynamic operation of the power grid. Historical plans and handling reports are mostly unstructured texts, relying on manual experience to extract and associate knowledge, lacking automated and high-precision entity and relationship extraction capabilities, leading to difficulties in knowledge accumulation and low reuse efficiency, failing to provide computationally viable knowledge support for intelligent decision-making. Finally, existing multi-agent systems lack effective knowledge-sharing mechanisms, and decisions made by various agents based on local information often conflict. Traditional methods rely on manual scheduling and multi-disciplinary coordination, resulting in slow response times and a lack of multi-objective collaborative optimization capabilities. Single models or centralized decision-making struggle to balance multiple constraints such as safety, stability, and economy, failing to achieve collaborative strategy generation among multiple agents. These problems result in existing systems generally exhibiting deficiencies such as large decision-making delays and low handling accuracy when dealing with complex scenarios such as cascading failures at new energy power plants and combined failures in AC / DC hybrid systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a power grid fault handling method and system based on multi-agent collaboration and knowledge graph-driven approaches. This addresses the problems of dispersed and heterogeneous multi-source fault information in the power grid, difficulty in assessing operational risks, unstructured power grid fault handling text, weak correlation of knowledge elements, difficulty in effectively supporting intelligent decision-making, low accuracy of contingency plan matching in power grid fault handling, and difficulties in multi-agent collaboration.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0006] This invention first discloses a power grid fault handling method based on multi-agent cooperation and knowledge graph-driven approach, which includes the following steps: Step 1: Merge multi-source power grid fault data, and perform real-time monitoring and fault risk analysis on the power grid operation status based on the fusion results, including cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis. Step 2: Based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, automatically identify and extract entity knowledge from the historical contingency plan text data of power grid fault handling and construct a knowledge graph of power grid fault handling; Step 3: Use the RE2 text similarity matching model to perform intelligent text matching of the fault risk analysis and the power grid fault handling knowledge graph, and use a multi-agent reinforcement learning algorithm based on the Actor-Critic framework to generate the optimal handling strategy for power grid faults.

[0007] The present invention further includes the following preferred embodiments: The fusion of multi-source power grid fault data further includes: Define a common data format that includes fields for timestamp, data source, measurement type, value, and device ID; parse and transform the raw data from each data source, align all the data using precise timestamps, and store them in a unified time-series database to form a single database; The key features that best characterize the state of each data source are extracted in parallel. Based on the unified high-precision time reference provided by the data layer fusion, the features extracted from different sources are precisely aligned within the fault time window. Combined with the power grid topology model, spatial correlation between features is established to form a comprehensive feature set containing time and space dimensions. The correlated high-dimensional feature vector is input into the feature fusion algorithm to remove redundant information and generate a unified feature vector.

[0008] The cross-section limit exceedance analysis includes real-time monitoring of the power grid operation status, dynamic statistics of newly emerging or more severely exceeded cross-sections after a fault occurs; when the real-time value of the active power flow of a cross-section exceeds its preset steady-state limit, it is determined that there is a risk of cross-section limit exceedance and a corresponding alarm is generated. The analysis of weak links in the power grid includes the use of N-1 and N-2 verification methods. N-1 considers single component failures, while N-2 considers multiple failures. By disconnecting lines or main transformers in the power grid topology model, the system scans for situations such as complete shutdown of power plants and substations, or the formation of dead or isolated power supply areas. The corresponding disconnected components and power plant information are recorded as weak links. In other words, by comparing the scan results before and after the fault, the system eliminates existing weak points in the base state and captures new changes in power grid vulnerability caused by the fault.

[0009] The BiLSTM-CRF deep learning model includes an embedding layer, a BiLSTM layer, and a CRF layer. First, the input text is segmented into words, and the segmentation results are converted into vector representations as input data for the model. Then, the BiLSTM layer performs bidirectional feature extraction on the text sequence while capturing contextual information. Finally, the CRF layer performs global sequence annotation on the features output by the BiLSTM to generate the final entity recognition result.

[0010] The TextCNN text convolutional neural network is further used for: For an input text sequence containing n words, each word is mapped to a k-dimensional word vector through an embedding layer, forming the input matrix:

[0011] in, For word vector matrix, This represents a vector concatenation operation; Feature extraction is performed using convolutional kernels of different sizes; for a convolutional kernel with width h... Its convolution operation is represented as:

[0012] in, Represents a local word window, It is the ReLU activation function. For bias terms; Features are extracted in parallel using convolutional kernels of various sizes, and max pooling is performed on each feature map:

[0013] Where h takes three sizes: 2, 3, and 5, and j represents the filter index; The pooled features are concatenated and then fed into a fully connected layer to complete the relation classification.

[0014]

[0015] Where y is the final probability distribution of relation categories.

[0016] The RE2 text similarity matching model combines the encoding layer and the alignment layer into a basic block. By stacking multiple such blocks and using residual connections between blocks, it leverages a multi-layered inter-sentence alignment mechanism to deeply capture the semantic matching relationship between two texts. The processing steps for each block are as follows: (1) Extract the word embedding representations of the two texts respectively, i.e. the original word sense vectors; if the current block is the first layer, directly input the original word sense vectors into the encoding layer; otherwise, concatenate the output of the previous block with the original word sense vectors and then input them into the encoding layer; (2) The encoding layer outputs a context vector, which is then concatenated with the input and output of the encoding layer into a long matrix via residual connection and input to the alignment layer; (3) In the alignment layer, the model generates an alignment vector containing contextual semantic features and passes the input and output of this layer to the comparison layer; (4) The comparison layer performs three operations: concatenation, subtraction and multiplication, to generate a vector containing matching features. If the current block is the last layer, the vector is output as the final matching result; otherwise, it is passed to the next block through the residual connection as the input of its encoding layer.

[0017] The multi-agent reinforcement learning algorithm based on the Actor-Critic framework further includes: During training, a centralized approach is adopted, with each agent having an independent Critic network to collect state and action information from all agents, thereby constructing a joint observation space and action space. Each Critic network trains the agent's Actor network based on global information. During the testing and deployment phases, each agent makes decisions based solely on its own state, without relying on global information. Specifically, for a multi-agent system, the set of states consisting of N agents is represented as follows: The action set is represented as The reward set is represented as The state input and action output of each agent are represented as follows: and ,in This represents the parameters of the Actor network; after the Actor network of an agent outputs an action based on its own state, the Critic network evaluates the state and action information of all agents, generating two centralized action value functions. ; During each interaction with the environment, the multi-agent system collects the current state observations of all agents, joint actions, immediate rewards from environmental feedback, and the new state of the environment after the interaction, combining this information into a global experience tuple. ;in, and These represent the system states before and after the interaction. and These represent the action and reward of agent i, respectively; during the training phase, they are selected from the experience replay queue. Use global sample experience as training data; for the agent The network update process is as follows: First, the target policy network Generate motion and add noise :

[0018] in , These are random noise parameters; Calculate the two target Critic networks value:

[0019]

[0020] In the formula: Indicates the system's global observation state at the next moment; This represents the joint action generated by the target policy network of all agents in the next moment; and Representing intelligent agents respectively The parameters of the two target Critic networks; Calculate the TD target:

[0021] It is an intelligent agent The immediate reward obtained from the environment after performing an action; It is a discount factor; Update the parameters of the two Critic networks. :

[0022] Calculate the mathematical expectation of an empirical sample regarding state, action, reward, and next state; Delayed updates to the Actor network and the target network:

[0023] Indicates about observation and actions Calculate the mathematical expectation from empirical samples; Indicates the intelligent agent Actor network parameters Calculate the gradient; Indicates the intelligent agent action Calculate the gradient; The target network is updated using a soft update method, which involves taking a weighted average of the parameters of the old target network and the parameters of the new target network, and then assigning the weighted result to the target network.

[0024]

[0025] In the formula: These are the target Critic network parameters. These are the parameters of the Critic network. These are the parameters of the Actor network. These are the target Actor network parameters. It is the target network update rate.

[0026] This invention also discloses a power grid fault handling system based on multi-agent collaboration and knowledge graph-driven methods, utilizing the aforementioned power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approaches, comprising: The fault risk analysis module is used to fuse multi-source power grid fault data and perform real-time monitoring and fault risk analysis of the power grid operation status based on the fusion results. It includes cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis. The entity recognition and extraction module is used to automatically identify and extract entity knowledge from historical contingency plan text data of power grid fault handling based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, and to construct a power grid fault handling knowledge graph. The disposal strategy generation module is used to perform intelligent text matching of the fault risk analysis and the power grid fault disposal knowledge graph using the RE2 text similarity matching model, and to generate the optimal disposal strategy for power grid faults using a multi-agent reinforcement learning algorithm based on the Actor-Critic framework.

[0027] Accordingly, this application also discloses a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the aforementioned power grid fault handling method based on multi-agent collaboration and knowledge graph driving.

[0028] Accordingly, this application also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned power grid fault handling method based on multi-agent collaboration and knowledge graph driving.

[0029] The beneficial effects of this invention are as follows: Compared with the prior art, this invention provides a power grid fault handling method and system based on multi-agent collaboration and knowledge graph-driven approach. Through multi-source fault data fusion and risk analysis, it achieves real-time monitoring and risk assessment of power grid operation status; it employs entity recognition based on BiLSTM-CRF and relation extraction technology based on TextCNN to construct a structured power grid fault handling knowledge graph from unstructured text; it achieves intelligent semantic matching of fault information and contingency plans through the RE2 model, and generates multi-agent collaborative optimization handling strategies based on the MATD3 algorithm, thereby realizing a closed-loop process from fault perception, knowledge construction to intelligent decision-making. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the BiLSTM-CRF model structure in this invention.

[0031] Figure 2 This is a schematic diagram of the TextCNN network structure in this invention.

[0032] Figure 3 This is a schematic diagram of the overall structure of the RE2 model method in this invention.

[0033] Figure 4 This is a schematic diagram of the multi-agent reinforcement learning algorithm framework in this invention.

[0034] Figure 5 This is a schematic diagram of the MATD3 algorithm update framework in this invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0037] To address the shortcomings of existing technologies, this invention proposes a power grid fault handling method and system based on multi-agent collaboration and knowledge graph-driven approaches. It also proposes a power grid operation status monitoring and assessment scheme that integrates multi-source perception and risk analysis. High-precision entity recognition for contingency plans is achieved through BiLSTM-CRF, and knowledge relationship extraction is performed using TextCNN, thereby constructing a structured and reasonable knowledge graph for power grid fault handling, providing reliable knowledge support for subsequent intelligent decision-making. High-precision semantic matching is achieved through the RE2 model to accurately retrieve applicable contingency plans, and a multi-agent collaboration mechanism is built based on the MATD3 algorithm to generate globally optimized fault handling strategies, significantly improving decision-making efficiency and system collaboration capabilities.

[0038] The power grid fault handling method disclosed in this invention, based on multi-agent cooperation and knowledge graph-driven approach, includes the following steps: Step 1: Merge multi-source power grid fault data, and perform real-time monitoring and fault risk analysis of power grid operation status based on the fusion results, including cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis.

[0039] The power grid dispatch and control system is a core component of the smart grid, responsible for the management of the entire process of power production, transmission, and distribution. It ensures the safe and reliable operation of the power grid by maintaining supply and demand balance and coordinating various grid components. Once a system failure or anomaly occurs, it can severely impact the stability of the power system. Therefore, extracting effective features from massive system logs and constructing a unified structured data representation, and integrating multi-source heterogeneous data through unified modeling to achieve highly accurate and timely fault detection, has become an important research direction for smart grid dispatch and control systems.

[0040] Multi-source fault data fusion comprehensively utilizes various types of information, including data sources from Energy Management Systems (EMS), Supervisory Data Acquisition and Monitoring Systems (SCADA), Advanced Measurement Systems (AMI), and Wide Area Measurement Systems (WAMS), aiming to leverage the strengths of each information source and achieve complementarity and synergy. Multi-source fault data fusion can be divided into data layer fusion and feature layer fusion.

[0041] Data layer fusion directly processes and merges raw data from different data sources. It constructs a unified data model centered on high-precision timestamps and establishes mapping and transformation rules for each data source, integrating real-time operational data from the SCADA system, fault-time electrical quantity data from fault recorders, and synchronization phasor data from the WAMS system into a unified storage format. The specific implementation process is as follows: First, a general data format is defined, including standard fields such as timestamp, data source, measurement type, value, and device ID. Then, the raw data from each data source is parsed and transformed. Finally, all data is aligned using precise timestamps and stored in a unified time-series database, forming a single database that provides comprehensive support for subsequent analysis. It fully utilizes the detailed information of the raw data, avoiding potential errors in subsequent processes and improving data accuracy. By fusing raw current and voltage waveform data, the start time and type of faults can be analyzed more accurately, providing a more reliable basis for fault analysis. Furthermore, data layer fusion effectively reduces data transmission volume, lowers bandwidth requirements, and improves transmission efficiency.

[0042] Building upon data layer fusion, feature layer fusion follows a systematic process that deeply integrates features extracted from different data sources. This process comprises three core steps: 1) Targeted Feature Extraction: Extract the key features that best characterize the state from various data sources in parallel. For example, extract the start-up, output, and circuit breaker opening / closing action sequences and precise time stamps of protection devices from the alarm and event sequence records of the SCADA system; extract in-depth electrical features such as fault phase, current and voltage transient steady-state amplitude, harmonic content, and impedance changes from the waveform data of the fault recorder using signal processing algorithms; and extract dynamic response features such as phase angle difference between the two ends of the critical line, system frequency change rate, and power oscillation amplitude from the synchronous phasor data of the WAMS system.

[0043] 2) Spatiotemporal Feature Correlation: Based on the unified high-precision time benchmark provided by data layer fusion, the features extracted from different sources are precisely aligned within the fault time window. Simultaneously, combined with the power grid topology model, spatial correlations between features are established, forming a comprehensive feature set encompassing both temporal and spatial dimensions.

[0044] 3) Fusion and Dimensionality Reduction: The associated high-dimensional feature vector is input into the feature fusion algorithm to remove redundant information and generate a unified feature vector that can comprehensively and concisely reflect the essence of the fault.

[0045] Through the complete fusion process described above, the data dimensionality is significantly reduced, simplifying the complexity of subsequent data analysis. For example, when processing fault recorder data, the key features extracted through fusion can represent fault information more efficiently and concisely. Furthermore, fusing features from different data sources can comprehensively reflect the essential characteristics of the fault. For instance, combining the switching action sequences provided by the SCADA system with the transient electrical features provided by the fault recorder can more accurately determine the scope, type, and evolution of the fault, thereby greatly improving the accuracy and reliability of fault handling decisions.

[0046] Power grid fault risk analysis is a core component of ensuring the safe and stable operation of the power grid, encompassing multiple key dimensions such as cross-section overload analysis, equipment overload analysis, analysis of weak links in the power grid, and analysis of power outages of main transformers or busbars.

[0047] Regarding cross-section limit exceedance analysis, the system monitors the power grid's operating status in real time. Within two minutes of a fault occurring, it dynamically identifies newly emerging or increasingly severe cross-sections. When the real-time value of the active power flow at a certain cross-section exceeds its preset steady-state limit, the system determines that there is a risk of cross-section limit exceedance and generates a corresponding alarm.

[0048] In terms of equipment overload analysis, the system scans the operating conditions of all main transformers and lines in real time. The judgment criteria for main transformer overload are: if the real-time value of active power exceeds its rated capacity, it is determined that there is a risk of main transformer overload; line overload is judged by comparing the real-time phase current with the upper limit value of line current. Once the real-time value exceeds the upper limit, it means that there is a risk of line overload.

[0049] In analyzing weak links in the power grid, two typical scenarios are identified: first, a fault-induced new single-line feeder (single-line grid connection), the power outage impact of which is usually limited to one or two substations; second, a fault-induced new local single-line grid connection, the power outage of which may affect multiple substations in the entire area. The specific analysis employs N-1 and N-2 verification methods: by disconnecting lines or main transformers in the power grid topology model (N-1 considers single component faults, N-2 considers multiple faults such as parallel connection on the same pole), the system scans for situations such as complete substation outages, dead zones or isolated power supply areas, and records the corresponding disconnected components and substation information as weak links. Weak link identification emphasizes new characteristics; that is, by comparing the scan results before and after the fault, pre-existing weak points are excluded, and the newly caused changes in power grid vulnerability due to the fault are accurately captured.

[0050] In terms of main transformer or busbar power outage analysis, this analysis aims to diagnose in real time whether critical nodes have lost power supply capability after a fault, and is the core of determining the direct impact range of the fault. The analysis process includes: First, based on the real-time topology and electrical quantities of the power grid, the power outage status is determined. When all circuit breakers connected to each side of the main transformer are in the open state and the electrical quantity measurement is zero, it is determined that the main transformer is out of power; when all incoming circuit breakers connected to a certain busbar are open and the busbar voltage is zero, it is determined that the busbar is out of power. Subsequently, the system will automatically analyze the impact range of the power outage, determine the downstream loads and substations affected by the power outage, and generate different levels of risk alarms based on the importance of the out-of-power components and the impact range, providing direct basis for dispatchers to quickly locate the core area of ​​the fault and formulate recovery strategies. This analysis complements the analysis of weak links in the power grid, together forming a static and dynamic assessment system for the structural risks of the power grid.

[0051] Through the above multi-dimensional and real-time risk analysis, the system can comprehensively and quickly identify potential risks in power grid operation, providing key support for dispatchers' fault handling decisions and effectively improving the reliability and resilience of power grid operation.

[0052] Step 2: Based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, automatically identify and extract entity knowledge from the historical contingency plan text data of power grid fault handling and construct a knowledge graph of power grid fault handling.

[0053] In terms of entity recognition in contingency plans, this model addresses the large amount of unstructured contingency plan text data in the power grid sector, including historical fault handling reports, emergency operating procedures, and dispatch execution documents. It utilizes a bidirectional long short-term memory network-conditional random field (BiLSTM-CRF) for parsing, enabling accurate identification and extraction of key entity elements from the text, such as power equipment, operational actions, and operating status. This model alleviates the gradient vanishing and gradient exploding problems commonly encountered during long sequence training in contingency plan texts, thus more accurately capturing entity information within the sequence and achieving efficient extraction of contingency plan entity knowledge from the text.

[0054] Specifically, the BiLSTM-CRF model includes an embedding layer, a BiLSTM layer, and a CRF layer. The specific process is as follows: First, the input text is segmented into words, and the segmentation results are converted into vector representations, which serve as the input data for the model. Then, the BiLSTM layer performs bidirectional feature extraction on the text sequence, which can simultaneously capture contextual information and improve the accuracy of entity recognition. Finally, the CRF layer performs global sequence annotation on the features output by the BiLSTM to generate the final entity recognition result.

[0055] By organically combining BiLSTM and CRF, this model can not only make full use of sequence context information, but also ensure the rationality of the predicted label sequence, and achieve accurate identification of the entities in the plan text.

[0056] The embedding layer is responsible for mapping words in the input text to word vectors. First, the labeled text dataset is preprocessed, including removing redundant special characters. Then, a layer with dimension is initialized. The matrix, where Represents the length of the text. This represents the dimension of the word vector. The dimensional vector matrix will serve as the input to the model for further processing in subsequent layers.

[0057] In the task of named entity recognition, entity recognition relies not only on information from past moments but also on information from future moments. Therefore, model training requires simultaneous access to past and future input features. To this end, this invention employs a Bidirectional Long Short-Term Memory (BiLSTM) network as the deep neural network model. The BiLSTM consists of two layers: a forward LSTM and a backward LSTM. The forward and backward LSTMs process the input data sequentially according to the time series, with the forward LSTM capturing past information and the backward LSTM capturing future information. The hidden states at time t are defined as follows: and ,in Forward LSTM, For backward LSTM, the specific calculation formula is shown below:

[0058]

[0059]

[0060] in, Let represent the input vector at time t; and These represent the hidden states of the forward and backward LSTM at time t, respectively. and They represent t respectively The hidden states of the forward and backward LSTM at time 1; and These are the hidden states in the forward and backward LSTM, respectively, corresponding to the hidden state in the previous time step. The weight matrix; and In the forward and backward LSTMs, respectively, the input corresponds to the current time step. The weight matrix; and These are the bias vectors for the forward and backward LSTMs, respectively. Indicates the activation function; This is a vector concatenation operation; It is a bidirectional LSTM.

[0061] The word vector sequence is input into the BiLSTM layer, and the forward LSTM obtains the forward hidden layer sequence of word vectors. It captures past contextual information. The backward LSTM then obtains the backward hidden layer sequence of word vectors. This captures future contextual information. By concatenating the hidden layer sequences of the forward and backward LSTM at each time step, a complete contextual representation is obtained, and the final hidden layer sequence is... And use it as input for the next layer.

[0062] Relying solely on BiLSTM layers can lead to ineffective output label sequences. To address this issue, this invention proposes connecting a Conditional Random Field (CRF) layer to the output of the BiLSTM. Unlike labeling individual elements, the CRF layer covers the entire sentence structure in sequence labeling tasks, optimizing the output label sequence by considering global dependencies. By using the output features generated by the BiLSTM layer as input to the CRF layer, the model can learn and apply specific constraints, thereby achieving more accurate sequence classification and labeling.

[0063] Let the input sequence be The probability matrix output by the BiLSTM layer is , express Marked as the The probability of each label. Then it represents the nth probability transition matrix. The label was transferred to the first... Each label probability.

[0064] For the output label sequence corresponding to the input sequence The path score formula is as follows. The score is calculated and the best label sequence is output.

[0065]

[0066]

[0067] In terms of entity relationship extraction from contingency plans, the semantic associations between different entities are determined based on their contextual semantics, syntactic structure, and logical features within the original contingency plan document. This invention constructs a contingency plan knowledge relationship extraction model based on TextCNN, thereby forming a structured knowledge system with rich semantics, providing computable and reasonable knowledge support for subsequent fault handling decisions. This model ensures the integrity of the extracted semantic relationships between entities and achieves efficient parsing of power grid fault handling contingency plan texts through the organic combination of entity recognition and entity relationship extraction technologies.

[0068] The TextCNN model uses convolutional windows of different sizes to capture local information within a sentence. These windows cover different segments of the sentence, extracting relevant features to achieve the extraction of entity relationships. Structurally, the network structure of the TextCNN model is as follows: Figure 2 As shown.

[0069] First, the contingency plan text is preprocessed and vectorized. Suppose the input text sequence contains n words; each word is mapped to a k-dimensional word vector through an embedding layer, forming the input matrix:

[0070] in, For word vector matrix, This indicates a vector concatenation operation.

[0071] Next, multi-scale convolutional feature extraction is performed. Feature extraction is conducted using convolutional kernels of different sizes; for a convolutional kernel with a width of h... Its convolution operation is represented as:

[0072] in, Represents a local word window, It is the ReLU activation function. This is a bias term.

[0073] Then, feature fusion and dimensionality reduction are performed. Features are extracted in parallel using convolutional kernels of various sizes, and max pooling is performed on each feature map.

[0074] Where h takes three sizes: 2, 3, and 5, and j represents the filter index.

[0075] Finally, relation classification is performed using a fully connected layer. The pooled features are concatenated and then fed into the fully connected layer.

[0076]

[0077] Where y is the final probability distribution of the relation categories. This is the weight matrix of the fully connected layer. This is the bias vector for the fully connected layer.

[0078] Through the above steps, this invention achieves efficient parsing of power grid fault handling plan texts, providing reliable knowledge support for fault handling decisions.

[0079] Step 3: Use the RE2 text similarity matching model to perform intelligent text matching of the fault risk analysis and the power grid fault handling knowledge graph, and use a multi-agent reinforcement learning algorithm based on the Actor-Critic framework to generate the optimal handling strategy for power grid faults.

[0080] The RE2 model is trained on the contingency plan text by fusing residual vectors, word embedding vectors, and encoding vectors to form an efficient semantic matching architecture. When fault alarm information is pushed to the fault handling contingency plan knowledge graph, the RE2 text similarity matching model is used to infer and match the alarm information with the fault handling trigger condition nodes in the knowledge graph. This model significantly reduces the parameter scale and computational complexity, while improving the accuracy and efficiency of matching, achieving rapid mapping from alarm information to corresponding handling nodes in the contingency plan knowledge graph. Based on the matched trigger condition nodes, the corresponding contingency plan handling module is located, all related handling method nodes are obtained, and unified push is performed, ultimately completing the intelligent matching and recommendation of fault handling.

[0081] The overall structure of the RE2 model is as follows: Figure 3 As shown. This model combines the encoding layer and alignment layer into a basic block. By stacking multiple such blocks and using residual connections between blocks, it leverages a multi-layered inter-sentence alignment mechanism to deeply capture the semantic matching relationship between two texts. The processing steps for each block are as follows: (1) Extract the word embedding representations of the two texts respectively, that is, the original word meaning vectors (corresponding to Figure 3 (The blank matrix in the code). If the current block is the first layer, the original word sense vector is directly input into the encoding layer; otherwise, the output of the previous block is concatenated with the original word sense vector before being input into the encoding layer.

[0082] (2) The encoding layer outputs the context vector ( Figure 3 The black matrix in the image is used to concatenate the input and output of the coding layer into a long matrix through residual connections, and then input it to the alignment layer.

[0083] (3) In the alignment layer, the model generates an alignment vector containing contextual semantic features and passes the input and output of this layer to the comparison layer.

[0084] (4) The comparison layer performs three operations: concatenation, subtraction, and multiplication to generate a vector containing matching features. If the current block is the last layer, the vector is output as the final matching result; otherwise, it is passed to the next block through the residual connection as the input to its encoding layer.

[0085] To achieve the automatic generation of optimal handling strategies from real-time grid fault status, this invention models the fault handling decision-making process as a multi-agent cooperative optimization problem. Specifically, key control units in the power grid are abstracted as independent agents, such as generator controllers, circuit breaker controllers, and load switch controllers. Each agent makes decisions based on its local observations, including operational data such as node voltage and line power flow, and its decision actions encompass control commands such as adjusting output and switching states. The goal is to generate a globally optimal handling strategy through collaborative efforts that simultaneously optimizes multiple objectives, including grid security, power supply reliability, equipment risk, and economic efficiency.

[0086] This method is implemented based on a multi-agent deep reinforcement learning framework with centralized training and distributed execution. Its core principles and application process are as follows: During the training phase, the system constructs a centralized Critic network, which can acquire the state and action information of all agents, thereby learning to evaluate the global value of joint actions. This design enables agents to clearly understand the impact of their actions on the overall system during training, effectively addressing the non-stationarity problem in multi-agent environments. During the policy execution phase, each agent makes independent decisions based solely on its own Actor network and local observations, without requiring global information interaction, thus ensuring the real-time performance and reliability of the system's decisions in a real power grid environment.

[0087] In this invention, the framework is specifically applied as follows: the real-time fault state output by the data fusion and risk analysis module is mapped to the joint state space of a multi-agent system. Typical states include operational anomalies such as line L1 overload of 150% and bus B1 voltage exceeding limits. Each agent's Actor network outputs control actions based on this state. The actions of all agents collectively constitute a joint handling scheme, which is applied to a high-fidelity power grid simulation environment. The environment calculates a new operating state based on the power grid physical model and generates a comprehensive reward signal, such as a positive reward for eliminating limits and a negative reward for unnecessary load shedding. Through training with a large amount of such interactive experience, the Critic network gradually learns to accurately evaluate the long-term value of different handling schemes, thereby guiding the Actor networks of each agent to collaboratively optimize their strategies, and ultimately automatically generating a fault handling scheme that approximates the global optimum.

[0088] Multi-agent systems consist of multiple individuals with certain sensing, computing, and execution capabilities, communicating and collaborating through a network. This reduces the complexity of individual agents while effectively improving the robustness, reliability, and flexibility of the entire system. Deep reinforcement learning combines the advantages of deep learning and reinforcement learning, achieving end-to-end learning from raw input to decision-making actions. Applying these ideas and algorithms to the learning and control of multi-agent systems is an important method for developing multi-agent systems with swarm intelligence.

[0089] Centralized training and distributed execution is a widely used multi-agent reinforcement learning architecture. (See also...) Figure 4 Based on the Actor-Critic framework, the Multi-Agent Dual-Delay Deep Deterministic Policy Gradient (MATD3) algorithm introduces a centralized value network to facilitate collaborative decision-making among multiple agents. During training, a centralized approach is employed, with each agent having an independent Critic network to collect state and action information from all agents, thus constructing a more complete joint observation and action space. This means that each Critic network can train the agent's Actor network based on global information. During testing and deployment, each agent makes decisions solely based on its own state, without relying on global information. This allows agents to make independent decisions in dynamically changing or unknown environments, thereby improving the system's scalability.

[0090] Specifically, for a multi-agent system, the set of states consisting of N agents is represented as follows: The action set is represented as The reward set is represented as The state input and action output of each agent are represented as follows: and ,in This represents the parameters of the Actor network. After the Actor network of an agent outputs an action based on its own state, the Critic network evaluates the state and action information of all agents, generating two centralized action value functions. This input encompasses the state and action information of all agents, resolving the non-stationary problem caused by changes in the behavior of other agents in the environment.

[0091] During system operation, multiple agents collect data through continuous interaction with the environment. At each interaction, the system collects the current state observations of all agents, their joint actions, the immediate rewards from the environment, and the new state of the environment after the interaction. This information is combined into a global experience tuple, denoted as […]. .in, and These represent the system states before and after the interaction. and These represent the action and reward of agent i, respectively. During operation, the system continuously stores the generated experience tuples into an experience replay queue, from which batches of data are randomly sampled during training to update the network parameters.

[0092] During the training phase, select from the experience replay queue The global sample experience is used as training data. For the agent... The network update process is as follows: First, the target policy network Generate motion and add noise :

[0093] in , This represents random noise parameters.

[0094] Calculate the two target Critic networks value:

[0095]

[0096] In the formula: Indicates the system's global observation state at the next moment; This represents the joint action generated by the target policy network of all agents in the next moment; and Representing intelligent agents respectively The parameters of the two target Critic networks.

[0097] Calculate the TD target:

[0098] In the formula: It is an intelligent agent The immediate reward obtained from the environment after performing an action; It is a discount factor.

[0099] Update the parameters of the two Critic networks. :

[0100] In the formula: Calculate the mathematical expectation of an empirical sample of states, actions, rewards, and the next state.

[0101] Delayed updates to the Actor network and the target network:

[0102] In the formula: Indicates about observation and actions Calculate the mathematical expectation from empirical samples; Indicates the intelligent agent Actor network parameters Calculate the gradient; Indicates the intelligent agent action Calculate the gradient.

[0103] The target network is updated using a soft update method, which involves taking a weighted average of the parameters of the old target network and the parameters of the new target network, and then assigning the weighted result to the target network.

[0104]

[0105] In the formula: These are the target Critic network parameters. These are the parameters of the Critic network. These are the parameters of the Actor network. These are the target Actor network parameters. It is the target network update rate.

[0106] The MATD3 algorithm update framework is as follows: Figure 5 As shown.

[0107] The beneficial effects of this invention are as follows: Compared with the prior art, this invention provides a power grid fault handling method and system based on multi-agent collaboration and knowledge graph-driven approach. Through multi-source fault data fusion and risk analysis, it achieves real-time monitoring and risk assessment of power grid operation status; it employs entity recognition based on BiLSTM-CRF and relation extraction technology based on TextCNN to construct a structured power grid fault handling knowledge graph from unstructured text; it achieves intelligent semantic matching of fault information and contingency plans through the RE2 model, and generates multi-agent collaborative optimization handling strategies based on the MATD3 algorithm, thereby realizing a closed-loop process from fault perception, knowledge construction to intelligent decision-making.

[0108] This invention can be a system, method, and / or computer program product. This invention also discloses a power grid fault handling system based on the aforementioned multi-agent collaboration and knowledge graph-driven power grid fault handling method, comprising: The fault risk analysis module is used to fuse multi-source power grid fault data and perform real-time monitoring and fault risk analysis of the power grid operation status based on the fusion results. It includes cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis. The entity recognition and extraction module is used to automatically identify and extract entity knowledge from historical contingency plan text data of power grid fault handling based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, and to construct a power grid fault handling knowledge graph. The disposal strategy generation module is used to perform intelligent text matching of the fault risk analysis and the power grid fault disposal knowledge graph using the RE2 text similarity matching model, and to generate the optimal disposal strategy for power grid faults using a multi-agent reinforcement learning algorithm based on the Actor-Critic framework.

[0109] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned power grid fault handling method driven by multi-agent collaboration and knowledge graphs. The computer program product may include a computer-readable storage medium on which computer-readable program instructions are loaded to enable a processor to implement various aspects of this disclosure. Specifically, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the aforementioned power grid fault handling method driven by multi-agent collaboration and knowledge graphs.

[0110] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0111] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0112] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach, characterized in that, Includes the following steps: Step 1: Merge multi-source power grid fault data, and perform real-time monitoring and fault risk analysis on the power grid operation status based on the fusion results, including cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis. Step 2: Based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, automatically identify and extract entity knowledge from the historical contingency plan text data of power grid fault handling and construct a knowledge graph of power grid fault handling; Step 3: Use the RE2 text similarity matching model to perform intelligent text matching of the fault risk analysis and the power grid fault handling knowledge graph, and use a multi-agent reinforcement learning algorithm based on the Actor-Critic framework to generate the optimal handling strategy for power grid faults; The fusion of multi-source power grid fault data further includes: Define a common data format that includes fields for timestamp, data source, measurement type, value, and device ID; parse and transform the raw data from each data source, align all the data using precise timestamps, and store them in a unified time-series database to form a single database; Key features are extracted from various data sources in parallel. Based on the unified high-precision time benchmark provided by data layer fusion, the features extracted from different data sources are precisely aligned within the fault time window. Combined with the power grid topology model, spatial correlation between features is established to form a comprehensive feature set containing time and space dimensions. The correlated high-dimensional feature vector is input into the feature fusion algorithm to remove redundant information and generate a unified feature vector.

2. The power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach according to claim 1, characterized in that, The cross-section limit exceedance analysis includes real-time monitoring of the power grid operation status, dynamic statistics of newly emerging or more severely exceeded cross-sections after a fault occurs; when the real-time value of the active power flow of a cross-section exceeds its preset steady-state limit, it is determined that there is a risk of cross-section limit exceedance and a corresponding alarm is generated. The analysis of weak links in the power grid includes the use of N-1 and N-2 verification methods. N-1 considers single component failures, while N-2 considers multiple failures. By disconnecting lines or main transformers in the power grid topology model, the system scans for situations such as complete shutdown of power plants and substations, or the formation of dead or isolated power supply areas. The corresponding disconnected components and power plant information are recorded as weak links. In other words, by comparing the scan results before and after the fault, the system eliminates existing weak points in the base state and captures new changes in power grid vulnerability caused by the fault.

3. The power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach according to claim 2, characterized in that, The BiLSTM-CRF deep learning model includes an embedding layer, a BiLSTM layer, and a CRF layer. First, the input text is segmented into words, and the segmentation results are converted into vector representations as input data for the model. Then, the BiLSTM layer performs bidirectional feature extraction on the text sequence while capturing contextual information. Finally, the CRF layer performs global sequence annotation on the features output by the BiLSTM to generate the final entity recognition result.

4. The power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach according to claim 3, characterized in that, The TextCNN text convolutional neural network is further used for: For an input text sequence containing n words, each word is mapped to a k-dimensional word vector through an embedding layer, forming the input matrix: in, For word vector matrix, This represents a vector concatenation operation; Feature extraction is performed using convolutional kernels of different sizes; for a convolutional kernel with width h... Its convolution operation is represented as: in, Represents a local word window, It is the ReLU activation function. For bias terms; Features are extracted in parallel using convolutional kernels of various sizes, and max pooling is performed on each feature map: Where h takes three sizes: 2, 3, and 5, and j represents the filter index; The pooled features are concatenated and then fed into a fully connected layer to complete the relation classification. Where y is the final probability distribution of relation categories.

5. The power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach according to claim 4, characterized in that, The RE2 text similarity matching model combines the encoding layer and the alignment layer into a basic block. By stacking multiple such blocks and using residual connections between blocks, it leverages a multi-layered inter-sentence alignment mechanism to deeply capture the semantic matching relationship between two texts. The processing steps for each block are as follows: (1) Extract the word embedding representations of the two texts respectively, i.e. the original word sense vectors; if the current block is the first layer, directly input the original word sense vectors into the encoding layer; otherwise, concatenate the output of the previous block with the original word sense vectors and then input them into the encoding layer; (2) The encoding layer outputs a context vector, which is then concatenated with the input and output of the encoding layer into a long matrix via residual connection and input to the alignment layer; (3) In the alignment layer, the model generates an alignment vector containing contextual semantic features and passes the input and output of this layer to the comparison layer; (4) The comparison layer performs three operations: concatenation, subtraction and multiplication, to generate a vector containing matching features. If the current block is the last layer, the vector is output as the final matching result; otherwise, it is passed to the next block through the residual connection as the input of its encoding layer.

6. The power grid fault handling method based on multi-agent collaboration and knowledge graph-driven approach according to claim 5, characterized in that, The multi-agent reinforcement learning algorithm based on the Actor-Critic framework further includes: During training, a centralized approach is adopted, with each agent having an independent Critic network to collect state and action information from all agents, thereby constructing a joint observation space and action space. Each Critic network trains the agent's Actor network based on global information. During the testing and deployment phases, each agent makes decisions based solely on its own state, without relying on global information. Specifically, for a multi-agent system, the set of states consisting of N agents is represented as follows: The action set is represented as The reward set is represented as The state input and action output of each agent are represented as follows: and ,in This represents the parameters of the Actor network; after the Actor network of an agent outputs an action based on its own state, the Critic network evaluates the state and action information of all agents, generating two centralized action value functions. ; During each interaction with the environment, the multi-agent system collects the current state observations of all agents, joint actions, immediate rewards from environmental feedback, and the new state of the environment after the interaction, combining this information into a global experience tuple. ;in, and These represent the system states before and after the interaction. and These represent the action and reward of agent i, respectively; during the training phase, they are selected from the experience replay queue. Use global sample experience as training data; for the agent The network update process is as follows: First, the target policy network Generate motion and add noise : in , These are random noise parameters; Calculate the two target Critic networks value: In the formula: Indicates the system's global observation state at the next moment; This represents the joint action generated by the target policy network of all agents in the next moment; and Representing intelligent agents respectively The parameters of the two target Critic networks; Calculate the TD target: It is an intelligent agent The immediate reward obtained from the environment after performing an action; It is a discount factor; Update the parameters of the two Critic networks. : Calculate the mathematical expectation of an empirical sample regarding state, action, reward, and next state; Delayed updates to the Actor network and the target network: Indicates about observation and actions Calculate the mathematical expectation from empirical samples; Indicates the intelligent agent Actor network parameters Calculate the gradient; Indicates the intelligent agent action Calculate the gradient; The target network is updated using a soft update method, which involves taking a weighted average of the parameters of the old target network and the parameters of the new target network, and then assigning the weighted result to the target network. In the formula: These are the target Critic network parameters. These are the parameters of the Critic network. These are the parameters of the Actor network. These are the target Actor network parameters. It is the target network update rate.

7. A power grid fault handling system based on multi-agent collaboration and knowledge graph-driven approach, characterized in that, include: The fault risk analysis module is used to fuse multi-source power grid fault data and perform real-time monitoring and fault risk analysis of the power grid operation status based on the fusion results. It includes cross-section overload analysis, equipment overload analysis, power grid weak link analysis, and main transformer or bus power failure analysis. The entity recognition and extraction module is used to automatically identify and extract entity knowledge from historical contingency plan text data of power grid fault handling based on the BiLSTM-CRF deep learning model and TextCNN text convolutional neural network, and to construct a power grid fault handling knowledge graph. The disposal strategy generation module is used to perform intelligent text matching of the fault risk analysis and the power grid fault disposal knowledge graph using the RE2 text similarity matching model, and to generate the optimal disposal strategy for power grid faults using a multi-agent reinforcement learning algorithm based on the Actor-Critic framework. The fault risk analysis module is further used for: Define a common data format that includes fields for timestamp, data source, measurement type, value, and device ID; parse and transform the raw data from each data source, align all the data using precise timestamps, and store them in a unified time-series database to form a single database; Extract key features from various data sources in parallel; Based on the unified high-precision time benchmark provided by data layer fusion, features extracted from different data sources are precisely aligned within the fault time window. Combined with the power grid topology model, spatial correlations between features are established to form a comprehensive feature set containing time and spatial dimensions. The correlated high-dimensional feature vectors are then input into the feature fusion algorithm to remove redundant information and generate a unified feature vector.

8. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the power grid fault handling method based on multi-agent collaboration and knowledge graph driving according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the power grid fault handling method based on multi-agent collaboration and knowledge graph driven by any one of claims 1-6.

Images