Electric power system node attribution and reasoning method, system, equipment and medium

Abstract

The invention relates to the technical field of power systems, and discloses a power system node attribution and reasoning method, system and device and a medium, and the method comprises the steps: obtaining first device data of a target power system; performing first preprocessing on the first equipment data, and establishing a time-varying causal graph model according to a first preprocessing result; and solving the time-varying causal graph model to obtain a decision strategy. According to the method, the accuracy and efficiency of power system safety monitoring are improved, and faults can be quickly positioned and repaired. Through feature extraction of preprocessing, key information is extracted to support data requirements of a time-varying causal graph model. The introduced time-varying causal graph model considers the time factor and the causal relationship, so that the operation state of the system is reflected more accurately. A decision strategy obtained by solving the model ensures safe operation of the power system, and power failure and economic loss caused by faults are reduced.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a power system node attribution and reasoning method, system, equipment and medium. Background Art

[0002] With the rapid development of smart grids and the deep integration of energy internet, the power system is gradually evolving towards digitalization and intelligence. As a key hub in the power cyber-physical system (CPS), the aggregation node undertakes functions such as data collection, transmission, aggregation and edge computing. However, with the large-scale access of Internet of Things (IoT) devices and the sophistication of network attack technologies, aggregation nodes face severe security threats, such as covert attacks (such as APT attacks), data tampering, and malicious code injection. Traditional security analysis methods (such as static detection based on rule engines or statistical learning models) are difficult to cope with dynamically changing network topologies and new attack modes. There is an urgent need for an intelligent method that can achieve real-time attribution and reasoning. Existing aggregation methods have the following limitations:

[0003] 1. Insufficient dynamic adaptability: Existing methods rely on fixed rules or offline training models and cannot adapt to dynamic changes in power system topology (such as node addition and removal, link reconstruction) and the rapid evolution of attack patterns.

[0004] 2. Difficulty in fusing multi-source heterogeneous data: The aggregation node needs to process multimodal data from sensors, SCADA systems, and communication protocols (such as IEC61850). Traditional methods find it difficult to effectively fuse spatiotemporal features with protocol semantics.

[0005] 3. Lack of causal reasoning capabilities: Most solutions are limited to correlation analysis and lack in-depth reasoning about attack paths and root causes, resulting in high false alarm rates and delayed response decisions.

[0006] 4. Real-time bottleneck: Existing deep learning-based models (such as LSTM and GCN) have high computational complexity and cannot meet the millisecond-level response requirements of the power system. Summary of the Invention

[0007] In view of the above existing problems, this application is proposed.

[0008] Therefore, the present application provides a power system node attribution and reasoning method, system, equipment and medium, which can solve problems such as insufficient dynamic adaptability, difficulty in fusing multi-source heterogeneous data, lack of causal reasoning ability and real-time bottleneck.

[0009] To solve the above technical problems, this application provides the following technical solutions:

[0010] In a first aspect, the present application provides a power system node attribution and reasoning method, including:

[0011] Acquiring data of a first device in a target power system;

[0012] The first device data includes real-time power parameter data, monitoring data and communication data;

[0013] performing a first preprocessing on the first device data, the first preprocessing comprising a feature extraction operation;

[0014] The feature extraction operation is achieved by establishing a feature extraction network;

[0015] Establishing a time-varying causal graph model based on the result of the first preprocessing;

[0016] The time-varying causal graph model is implemented by an improved first algorithm, wherein the first algorithm includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm;

[0017] The time-varying causal graph model is solved to obtain a decision strategy.

[0018] As a preferred solution of the power system node attribution and reasoning method described in this application, wherein: the establishing of a time-varying causal graph model based on the result of the first preprocessing includes:

[0019] Performing high-dimensional transformation on the result of the first preprocessing to obtain the result of the first preprocessing in a high-dimensional feature space;

[0020] Performing an improved causal analysis algorithm on the results of the first preprocessing to obtain a plurality of causal graphs;

[0021] Adjusting the plurality of causal graphs according to the improved Bayesian structure learning algorithm;

[0022] A time-varying causal graph model is established based on several adjusted causal graphs.

[0023] This preferred solution can more accurately capture the dynamic changes in the power system and the complex causal relationships between various devices. The high-dimensional transformation step makes the feature space richer, which helps to discover the associated information hidden deep in the data. The application of the improved causal analysis algorithm enhances the model's ability to identify causal relationships and reduces the possibility of misjudgment. The introduction of the improved Bayesian structural learning algorithm further optimizes the structure of the causal graph, making it more consistent with the actual situation of the power system. These improvement measures work together to make the established time-varying causal graph model more accurate and reliable, providing a solid foundation for the formulation of subsequent decision-making strategies.

[0024] As a preferred solution of the power system node attribution and reasoning method described in this application, wherein: the establishment of a time-varying causal graph model based on the adjusted causal graphs includes:

[0025] Calculating a causal weight matrix of each node according to the adjusted causal graphs;

[0026] performing an estimation operation on the causal weights in the causal weight matrix;

[0027] The causal weight matrix after the estimation operation is completed is used as the time-varying causal graph model.

[0028] As a preferred solution of the power system node attribution and reasoning method described in this application, wherein: the solving of the time-varying causal graph model to obtain the decision strategy includes:

[0029] Establish a scoring mechanism for evaluating candidate root cause nodes;

[0030] Solving the time-varying causal graph model according to the scoring mechanism;

[0031] Determine the decision-making strategy based on the solution results.

[0032] As a preferred solution of the power system node attribution and reasoning method described in this application, wherein: the establishment of the feature extraction network includes:

[0033] performing normalization processing on the first device data of the target power system;

[0034] Perform convolutional neural network feature extraction on the normalized data of the first device of the target power system;

[0035] The convolutional neural network feature extraction includes the introduction of the attention mechanism;

[0036] Establish a dynamic adjacency matrix and node attribute embedding results based on convolutional neural network feature extraction;

[0037] Feature fusion is performed according to the dynamic adjacency matrix and the node attribute embedding result to obtain a first preprocessing result.

[0038] As an optimal solution for the power system node attribution and reasoning method described in this application, the dynamic adjacency matrix is a matrix that displays the dynamic changes in the connection relationship between each node in the target power system under time steps, the nodes are each device in the target power system, and the matrix elements are the connection strength and status between each device.

[0039] As a preferred solution of the power system node attribution and reasoning method described in this application, wherein: the node attribute embedding result is the result of mapping the node attribute information into a low-dimensional vector space;

[0040] The attribute information of the node includes device type and rated power.

[0041] In a second aspect, the present application provides a power system node attribution and reasoning system, including:

[0042] A data acquisition module, configured to acquire data of a first device in a target power system;

[0043] The first device data includes real-time power parameter data, monitoring data and communication data;

[0044] a preprocessing module, configured to perform a first preprocessing on the first device data, wherein the first preprocessing includes a feature extraction operation;

[0045] The feature extraction operation is achieved by establishing a feature extraction network;

[0046] A model building module, configured to build a time-varying causal graph model based on the result of the first preprocessing;

[0047] The time-varying causal graph model is implemented by an improved first algorithm, wherein the first algorithm includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm;

[0048] A solution module is used to solve the time-varying causal graph model to obtain a decision strategy.

[0049] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described above when executing the computer program.

[0050] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described above when the computer program is executed by a processor.

[0051] Compared with existing technologies, the present application has the following beneficial effects: This application proposes a power system node attribution and reasoning method, which obtains first device data of a target power system; performs a first preprocessing on the first device data, which includes a feature extraction operation; establishes a time-varying causal graph model based on the results of the first preprocessing; and solves the time-varying causal graph model to obtain a decision strategy. This method not only improves the accuracy and efficiency of power system security monitoring but also enables the rapid location and repair of power system faults. Specifically, the feature extraction operation in the first preprocessing can effectively extract key information related to the target power system security, providing data support for the subsequent establishment of the time-varying causal graph model. Furthermore, the introduction of the time-varying causal graph model enables the present application to fully consider the time factor and causal relationships in the power system, thereby more accurately reflecting the actual operating status of the power system. Furthermore, by solving the time-varying causal graph model, the resulting decision strategy can provide strong guarantees for the safe operation of the power system and reduce power outages and economic losses caused by faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 A flowchart of a method for attributing and reasoning nodes in a power system provided in one embodiment of the present application.

[0054] Figure 2 An internal structure diagram of an electronic device for a power system node attribution and reasoning method provided in one embodiment of the present application. DETAILED DESCRIPTION

[0055] To make the above-mentioned purposes, features, and advantages of this application more clearly understood, the following detailed description of the specific embodiments of this application is given in conjunction with the accompanying drawings. It is obvious that the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of this application.

[0056] Example 1, with reference to Figure 1-Figure 2 , which is the first embodiment of the present application, provides a power system node attribution and reasoning method, including:

[0057] Before introducing the embodiments of the present application in detail, some related concepts are first explained for the sake of clarity.

[0058] Dynamic Causal Network (DCN): A network model used to analyze and process the dynamic changes in causal relationships in a system over time. In this application, it combines dynamic topology awareness technology with causal strength quantification methods to monitor changes in power system topology in real time and accurately quantify the causal influence strength between various factors. Compared to traditional causal analysis models, DCN can better adapt to the dynamic characteristics of power systems and provide a more accurate and real-time causal relationship model foundation for security attribution and reasoning.

[0059] Parameterized quantization technology is a technique for processing model parameters during model training and updating. By mapping continuous parameter values to discrete quantization levels, the space required for model storage and computational complexity are reduced while ensuring model performance. In this application's online incremental learning mechanism, parameterized quantization technology is utilized to rapidly update model parameters, reducing model update latency to less than 50ms, meeting the strict real-time requirements of power systems.

[0060] Existing technologies present several challenges. For example, traditional attribution and reasoning methods for power system fault events often rely on manual analysis, which is not only inefficient but also susceptible to human factors, resulting in low accuracy and reliability of analysis results. Furthermore, as power systems continue to expand and become more complex, the diversity and complexity of fault events are also increasing. Traditional attribution and reasoning methods are no longer able to meet the needs of practical applications.

[0061] This application provides a method that can effectively solve the above-mentioned problems. Next, we will explain in detail how to implement the power system node attribution and reasoning method in combination with multiple embodiments;

[0062] Figure 1 A flowchart of a method for attributing and reasoning nodes in a power system is shown, including:

[0063] S101, obtaining data of a first device in a target power system;

[0064] It should be noted that in order to realize power system node attribution and reasoning, this application needs to obtain relevant parameters about the target power system. In order to realize accurate security aggregation node attribution and reasoning as much as possible, more types of data need to be obtained.

[0065] In an optional embodiment, for example, it is necessary to obtain real-time equipment operating data, historical fault data, equipment configuration parameters, environmental factor data, etc. This data will be input into the parameterized quantization technology, and through efficient algorithm processing, key features will be extracted to construct an accurate model of the power system.

[0066] In another optional embodiment, machine learning algorithms are used to analyze and mine this data, identifying potential failure modes and patterns, providing strong support for subsequent attribution and reasoning. This approach can greatly improve the accuracy and efficiency of attribution and reasoning, providing a strong guarantee for the safe operation of the power system.

[0067] However, in the embodiment of the present application, the first device data includes real-time power parameter data, monitoring data, and communication data. These three different types of parameter data are sufficient to provide comprehensive and critical information for attribution and reasoning of power system nodes.

[0068] Among them, real-time power parameter data can reflect the current operating status of the power system, such as voltage, current, frequency, etc. These data are important bases for judging whether the power system is operating stably.

[0069] Among them, the monitoring data includes the equipment's temperature, humidity, vibration and other status information. Through these data, abnormal conditions of the equipment can be discovered in a timely manner to prevent potential safety hazards.

[0070] Among them, communication data records the information interaction between devices, which helps to analyze the coordination and communication failures between devices.

[0071] It should be noted that by analyzing these three types of data together, we can more accurately locate the problems of security aggregation nodes and improve the accuracy and efficiency of attribution and reasoning.

[0072] It should also be noted that in power system security analysis, multi-source heterogeneous data contains a wealth of information, but the modalities, structures, and characteristics of this data vary, making direct processing and analysis difficult. Therefore, it is necessary to fuse multimodal data such as real-time power parameter data, monitoring data, and communication data, and extract effective features to enable subsequent accurate safety attribution and reasoning.

[0073] S102, performing a first preprocessing on the first device data, where the first preprocessing includes a feature extraction operation;

[0074] It should be noted that because the acquired first device data is messy and the data has different modes, structures and characteristics, it is necessary to preprocess the first device data and use the preprocessed data to perform power system node attribution and reasoning.

[0075] In an optional embodiment, the first preprocessing step may include data cleaning operations, such as removing outliers and filling missing values, to ensure data integrity and accuracy. Furthermore, data dimensionality reduction can be performed to extract key features through methods such as principal component analysis, reducing data complexity and improving the efficiency of subsequent processing. These preprocessing steps provide a reliable data foundation for subsequent feature extraction, attribution, and inference.

[0076] However, ordinary preprocessing cannot meet the requirements of the subsequent establishment of a time-varying causal graph model in this application, so it is necessary to design a non-existing first preprocessing method.

[0077] In the embodiment of the present application, the feature extraction operation is achieved by establishing a feature extraction network.

[0078] In an optional embodiment, the feature extraction network can perform feature extraction on the preprocessed data using a deep learning algorithm, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). These networks can automatically learn and extract key features from complex data without the need for manual feature selection, thereby greatly improving the efficiency and accuracy of feature extraction.

[0079] In an optional embodiment, the feature extraction network can further improve feature extraction performance by combining the results of multiple base learners using an ensemble learning method, such as random forest or gradient boosting tree. This method can fully leverage the advantages of different base learners and improve the stability and robustness of feature extraction.

[0080] It should be noted that the benefits of using feature extraction networks for feature extraction are obvious.

[0081] First, it can automatically complete the feature extraction process, greatly reducing the manual workload.

[0082] Secondly, since the feature extraction network can automatically learn key features from the data, the extracted features are often more accurate and effective.

[0083] Finally, the feature extraction network can adapt to different types of data and has strong versatility and flexibility. These advantages make the feature extraction network an important tool for power system node attribution and reasoning.

[0084] In an embodiment of the present application, establishing a feature extraction network includes:

[0085] performing normalization processing on the first device data of the target power system;

[0086] Perform convolutional neural network feature extraction on the normalized data of the first device of the target power system;

[0087] Convolutional neural network feature extraction includes the introduction of attention mechanism;

[0088] Establish a dynamic adjacency matrix and node attribute embedding results based on convolutional neural network feature extraction;

[0089] Feature fusion is performed based on the dynamic adjacency matrix and the node attribute embedding results to obtain the first preprocessing result.

[0090] In an embodiment of the present application, the dynamic adjacency matrix is a matrix that displays the dynamic changes in the connection relationship between each node in the target power system at a time step. The nodes are each device in the target power system, and the matrix elements are the connection strength and status between each device.

[0091] In the embodiment of the present application, the node attribute embedding result is the result of mapping the node attribute information into a low-dimensional vector space;

[0092] The node's attribute information includes device type and rated power.

[0093] Specifically, multimodal data, including power data streams, equipment logs, and network traffic (i.e., real-time power parameter data, monitoring data, and communication data), is cleaned to remove noise and outliers. For example, sudden spikes in power data streams can be removed using filtering algorithms.

[0094] Furthermore, the data from different modalities are normalized to bring them into the same scale range, to avoid the impact of large scale differences on subsequent feature extraction and fusion. Common normalization methods include min-max normalization and Z-score normalization.

[0095] Furthermore, for each modality of data First, the data is fed into a CNN layer for local feature extraction. A CNN layer typically contains multiple convolution kernels, each capable of extracting different local features. Parameters such as kernel size, number, and stride need to be adjusted based on the specific data characteristics and task requirements.

[0096] Furthermore, the output of the CNN layer is fed into an LSTM layer, which learns the temporal dependencies of the data. The LSTM unit contains an input gate, a forget gate, and an output gate. These gating mechanisms allow it to effectively process long sequences of data.

[0097] Furthermore, the attention mechanism is used to calculate the attention weight α of each modality i Specifically, it can be achieved through a fully connected layer and a softmax function. First, The output of is input to the fully connected layer for linear transformation, and then the output is converted into a probability distribution through the softmax function to obtain the attention weight α i .

[0098] Furthermore, dynamic adjacency matrix and node attribute embedding processing, dynamic adjacency matrix A t It can be updated by monitoring the topology changes of the power system in real time. For example, when a new device is connected or disconnected, the value of the matrix element is updated accordingly.

[0099] Node attribute embedding E t This can be achieved through a pre-trained embedding model or autoencoder. The attribute information of the node is input into the embedding model, and the corresponding low-dimensional vector representation is output.

[0100] Furthermore, feature fusion is performed according to the feature fusion formula to obtain the final fusion feature vector F t .

[0101] Through the above steps, the fusion and feature encoding of multi-source data are achieved, providing high-quality feature input for subsequent causal analysis and security attribution reasoning.

[0102] In an optional embodiment, the feature fusion formula may be as follows:

[0103]

[0104] Among them: F t : represents the fused feature vector at time step t. This vector integrates the characteristic information of multimodal data and can more comprehensively and accurately reflect the state of the power system at that moment, providing a basis for subsequent analysis.

[0105] N: represents the number of modes of multimodal data.

[0106] In an optional embodiment, for example, power data flow, device logs, and network traffic are each considered a modality, so N = 3. Different modal data describe the operation of the power system from different perspectives. By fusing these modal data, richer information can be obtained.

[0107] α i : is the attention weight of each modality, and satisfies

[0108] In an optional embodiment, the attention weights reflect the importance of each modality in the fusion process, and the system automatically adjusts these weights based on the characteristics of the data and the current task requirements. For example, in some cases, the power data flow may have a greater impact on the system state, and the corresponding α i The value will be relatively large.

[0109] is the data of the i-th mode at time step t Perform feature extraction operations.

[0110] In an optional embodiment, a convolutional neural network (CNN) is used: CNNs have powerful local feature extraction capabilities, capable of capturing spatial characteristics within data. For data such as power data streams, equipment logs, and network traffic, CNNs can extract local patterns and features. For example, when processing power data streams, CNNs can identify local fluctuations in parameters such as voltage and current.

[0111] In an optional embodiment, the LSTM (Long Short-Term Memory) network is used: LSTM excels at processing sequential data and can capture temporal dependencies within the data. Power system data typically exhibits time series characteristics, and LSTM can learn how data changes over time. For example, it can predict changing trends in power load.

[0112] In summary, CNN-LSTM is able to simultaneously extract spatial and temporal features of multimodal data and convert them into representative feature vectors.

[0113] CNN-LSTM combination: In this application, the CNN-LSTM combination is not a simple application of existing technology. Targeted improvements have been made to it when processing multimodal data of the power system. For example, in terms of network structure and parameter settings, the size, number, step size of the CNN convolution kernel and the gating mechanism parameters of the LSTM unit are optimized according to the characteristics of data such as power data streams, equipment logs, and network traffic. At the same time, combined with the spatiotemporal attention mechanism, it can more effectively capture the spatiotemporal characteristics of power system data. This improved CNN-LSTM combination can better mine the power protocol semantics and physical constraint information contained in the data, providing more valuable feature input for subsequent safety attribution and reasoning.

[0114] β: This is a hyperparameter that controls the weight of the dynamic adjacency matrix and node attribute embedding fusion in the final feature vector. Its value range usually needs to be adjusted through experiments to achieve the best fusion effect.

[0115] CAT(A t ,E t ): Indicates that the dynamic adjacency matrix A t and node attribute embedding E t Perform splicing operations.

[0116] Dynamic adjacency matrix A t : Describes the dynamic changes in the connection relationship between nodes in the power system at time step t.

[0117] In an alternative embodiment, for example, in a power network, nodes can represent devices such as generators and substations, and matrix elements represent the connection strength and status between devices. The dynamic adjacency matrix can reflect changes in the topology of the power system.

[0118] Node attribute embedding E t : It is the result of mapping the node attribute information (such as device type, rated power, etc.) to a low-dimensional vector space.

[0119] In an optional embodiment, node attribute embedding can convert the complex attribute information of the node into a vector form that is easy for computers to process, facilitating subsequent calculations and analysis.

[0120] It should be noted that performing a first preprocessing on the first device data, including feature extraction, can improve the efficiency and accuracy of subsequent steps. Feature extraction can extract key information from the raw data, remove redundancy and noise, and make the data more streamlined and representative. This helps the model better capture dynamic changes in the power system and improve the accuracy of attribution and reasoning. Furthermore, the preprocessed data is formatted in a unified format, facilitating subsequent processing and analysis by algorithms and models, thereby accelerating the entire node attribution and reasoning process.

[0121] S103, establishing a time-varying causal graph model based on the result of the first preprocessing;

[0122] It should be noted that the purpose of establishing a time-varying causal graph model is to achieve node attribution and reasoning. The use of causal graphs can intuitively display the dynamic causal relationship between nodes and capture the time-varying characteristics and complex interactive relationships in the power system. By constructing a time-varying causal graph model, this application can more accurately understand the causal chain between different nodes and reveal the path of fault or anomaly propagation. This method helps to quickly locate the source of the problem, analyze the scope of impact, and take corresponding countermeasures when a security incident occurs in the power system.

[0123] In an optional embodiment, a time-varying causal graph model can be established using a dynamic Bayesian network approach based on time series data. This approach analyzes the statistical dependencies within time series data to construct a time-varying causal structure, accurately reflecting the dynamic causal relationships between power system nodes. Furthermore, by combining expert knowledge and historical data, model parameters can be further optimized, improving both accuracy and robustness.

[0124] In an optional embodiment, the time-varying causal graph model can also be established using deep learning methods. By designing a suitable neural network structure, the complex features and high-order dependencies in the input data are automatically learned to generate a time-varying causal graph. This method has powerful nonlinear modeling capabilities and can capture the complex dynamic characteristics of power systems. Furthermore, combined with transfer learning and incremental learning methods, the model's generalization and adaptability can be continuously improved.

[0125] In an embodiment of the present application, the time-varying causal graph model is performed by an improved first algorithm, which includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm;

[0126] In an embodiment of the present application, establishing a time-varying causal graph model according to the result of the first preprocessing includes:

[0127] Performing high-dimensional transformation on the result of the first preprocessing to obtain the result of the first preprocessing in a high-dimensional feature space;

[0128] Performing an improved causal analysis algorithm on the results of the first preprocessing to obtain several causal graphs;

[0129] Adjusting several causal graphs based on an improved Bayesian structure learning algorithm;

[0130] A time-varying causal graph model is established based on several adjusted causal graphs.

[0131] In an embodiment of the present application, establishing a time-varying causal graph model based on the adjusted causal graphs includes:

[0132] Calculate the causal weight matrix of each node based on the adjusted causal graphs;

[0133] Performing estimation operations on the causal weights in the causal weight matrix;

[0134] The causal weight matrix after the estimation operation is completed is used as the time-varying causal graph model.

[0135] Specifically, let’s first review the traditional Granger causality analysis: Traditional Granger causality analysis is used to determine whether one time series can help predict another time series.

[0136] Assume that this application has two time series X t and Y t If the t After the historical information of Y t The prediction accuracy of X is significantly improved. t Granger causality for Y t .

[0137] Improvement measures: In power systems, data often has complex nonlinear and dynamic characteristics, and the linear assumption of traditional Granger causality analysis is no longer applicable. Therefore, this application adopts a nonlinear Granger causality analysis method based on kernel functions as an improved causal analysis algorithm.

[0138] Specifically, by introducing a kernel function to map the original data into a high-dimensional feature space, linear Granger causality analysis is performed in the high-dimensional space. The kernel function can be a Gaussian kernel function:

[0139]

[0140] where x i and x j is a data point, and σ is the kernel bandwidth parameter.

[0141] Regarding improving the Bayesian structure learning algorithm, the main step is to adapt the traditional Bayesian structure learning algorithm to the present application to achieve specific improvements. The specific steps are as follows:

[0142] Data preprocessing: Data D from time 1 to time t 1:t Clean and normalize the data to remove noise and outliers to ensure data quality. Use the minimum-maximum normalization method to scale the data to the [0,1] interval. The formula is:

[0143]

[0144] Among them, x is the original data, x min and x max are the minimum and maximum values of the data respectively.

[0145] Furthermore, we can initialize the causal graph structure: we can randomly initialize a causal graph G0 or construct an initial causal graph based on prior knowledge. The structure of the initial causal graph will affect the subsequent learning process, but through Bayesian structure learning, the model will gradually adjust the causal graph structure to adapt to the data.

[0146] Furthermore, the causal graph structure is iteratively updated: the Markov Chain Monte Carlo (MCMC) method is used to sample and update the causal graph structure.

[0147] In each iteration, according to the current causal graph G t and data D 1:t , calculate the posterior probability P(G t |D 1:t ).

[0148] Then, a new candidate causal graph G is generated by making a minor modification to the causal graph structure (such as adding or deleting an edge). t+1, and calculate its corresponding posterior probability P(G t+1 |D 1:t ).

[0149] The Metropolis-Hastings criterion is used to decide whether to accept the new causal graph structure.

[0150] Furthermore, after each update of the causal graph structure, the causal weight matrix of each node is calculated based on the causal graph obtained by the improved Granger causal analysis results and the Bayesian structure learning.

[0151] In an optional embodiment, the least squares method or regularized regression method can be used to estimate the causal weight. For example, for node i, a linear regression model can be constructed where y i is the observation value of node i, x is the observation value of other nodes, and ∈ is the error term. By minimizing the sum of squares of the error terms and combining the L1 regularization term λ The causal weight matrix can be obtained estimated value.

[0152] Through the above steps, a time-varying causal graph model was successfully constructed based on improved Granger causal analysis and Bayesian structure learning, which can accurately capture the causal relationship between variables in the power system that changes over time.

[0153] In an optional embodiment, based on improved Granger causal analysis and Bayesian structure learning, the formula for constructing a time-varying causal graph model is as follows:

[0154]

[0155] Where: P(G t |D 1:t ): Indicates that given the data D from time 1 to time t 1:t Under the condition of time t, the causal graph G t The posterior probability of . This probability reflects the degree of confidence in the current causal graph structure based on historical data. The goal of this application is to find the causal graph structure that maximizes this posterior probability.

[0156] n: Represents the number of nodes in the causal graph. In power systems, nodes can represent various power equipment, monitoring points, or operating parameters. Each node corresponds to a variable, and the connections between them represent causal influences.

[0157] is the causal weight matrix of node i. This matrix describes the degree of causal influence of other nodes on node i at time t. Each element in the matrix represents the strength of a causal relationship, with positive values indicating positive causal relationships and negative values indicating negative causal relationships.

[0158] λ: Sparsity constraint coefficient. It controls the sparsity of the causal graph and prevents the model from becoming overly complex. A larger λ value causes more causal weights to approach zero, resulting in a sparser causal graph; a smaller λ value allows for more causal relationships. By adjusting the value of λ, a balance can be achieved between model complexity and fitting ability.

[0159] In an optional embodiment, its value is based on the following: when constructing a time-varying causal graph model, the sparsity constraint coefficient λ is used to control the sparsity of the causal graph and balance the model complexity and fitting ability. Its specific value is determined by the cross-validation method. The training data is divided into multiple subsets, different λ values are tried on different subsets, and the performance indicators of the model under each value are calculated (such as the ability of the causal graph structure to explain the observed data, measured by likelihood probability). The λ value that makes the model have the best comprehensive performance in cross-validation is selected as the final value. For example, starting from a series of candidate values such as 0.01, 0.1, 1, 10, etc., the model is evaluated in multiple cross-validations. If it is found that when λ = 0.1, the model can have the best explanatory power for the observed data while maintaining a reasonable complexity, then the λ value is determined to be 0.1.

[0160] is a matrix The L1 norm of , that is, the sum of the absolute values of all elements in the matrix. Using the L1 norm as a constraint can promote the sparsity of the causal weight matrix, making the model easier to interpret and analyze.

[0161] is an exponential penalty term that penalizes the sparsity of the causal weight matrix. It reduces the probability values corresponding to causal weight matrices with more non-zero elements, thereby encouraging the model to learn a simpler causal structure.

[0162] d k : represents the observation data at time k, which can be various monitoring data in the power system, such as voltage, current, power, etc.

[0163] P(d k |G k ): Indicates that in the causal graph G k Under the condition that the data d at time k is observed k The likelihood probability of . This probability describes the ability of the causal graph structure to explain the observed data. The larger the value, the more reasonable the causal graph structure is in explaining the observed data.

[0164] It should be noted that, based on the results of the first preprocessing, establishing a time-varying causal graph model can more accurately capture the dynamic change relationship in the power system. The time-varying causal graph model can be adjusted as the system state changes, reflecting the changes in causal relationships at different time points in the power system, which is crucial for understanding and predicting phenomena such as fault propagation and load fluctuations in the power system. In addition, by establishing a time-varying causal graph model, the present application can more effectively trace the fault source, that is, determine the root cause of the fault, thereby providing strong support for the safe operation and maintenance of the power system.

[0165] S104, solving the time-varying causal graph model to obtain a decision strategy.

[0166] In the embodiment of the present application, the time-varying causal graph model is solved to obtain a decision strategy including:

[0167] Establish a scoring mechanism for evaluating candidate root cause nodes;

[0168] Solve the time-varying causal graph model based on the scoring mechanism;

[0169] Determine the decision-making strategy based on the solution results.

[0170] It should be noted that the relevant operation steps of Monte Carlo Tree Search (MCTS) are as follows:

[0171] Initialization: Construct a causal graph G = (V, E), where V is the node set and E is the edge set. Use the root node of the causal graph as the initial node of MCTS and create an empty search tree T.

[0172] Simulation process:

[0173] a. Selection: Starting from the root node of the search tree, select a child node according to a certain strategy (such as the UCT algorithm) until a leaf node is reached. The selection formula of the UCT algorithm is:

[0174]

[0175] Where n is the current node, Q(n) is the cumulative reward of node n, N(n) is the number of times node n has been visited, p(n) is the parent node of node n, and c is an exploration constant.

[0176] b. Expansion: If the leaf node is not a terminal node, then expand one or more child nodes under the node. The expansion of child nodes can be performed based on the edge relationships in the causal graph, that is, starting from the current node, new nodes are generated along the edges in the causal graph.

[0177] c. Simulation: Starting from the expanded child node, a random simulation is performed to generate an attack propagation path. τDuring the simulation, the next node to be attacked is randomly selected according to the causal relationship and probability distribution in the causal graph until the termination condition is reached (such as reaching the final target node of the attack or reaching the maximum number of simulation steps). At the same time, the attack propagation delay Δt is recorded. τ .

[0178] d. Backtracking: Based on the simulation results, update the cumulative reward Q(n) and visit count N(n) for all nodes in the search tree from the root node to the simulated path. The cumulative reward can be calculated based on whether the simulated path contains the candidate root node and the attack propagation delay.

[0179] e. Termination condition: reaching the maximum number of simulation steps or reaching the final target node of the attack. The maximum number of simulation steps is set based on the complexity of the power system and computing resources. In the IEEE 39-node system test, the maximum number of simulation steps is set to 100. If the final target node of the attack is reached prematurely during the simulation, the simulation is terminated prematurely.

[0180] It should be noted that the relevant steps for applying transfer learning are as follows:

[0181] a. Knowledge transfer: First, collect causal relationship knowledge previously learned in similar power system fault diagnosis scenarios or on similar data. This knowledge can exist in the form of causal graph structures, model parameters, etc. For example, in previous fault diagnosis of power systems of similar scale, the causal relationships between certain generator nodes and substation nodes have been learned and extracted. In the new fault diagnosis scenario, the migrated causal relationship knowledge is adapted based on the structure and node information of the current causal graph. If the nodes in the new scenario have a corresponding relationship with the previous scenario, the relevant causal relationship is directly applied to the new causal graph; if there are some differences, appropriate adjustments are made based on the node attributes and system characteristics.

[0182] b. Model fine-tuning: During the Monte Carlo Tree Search (MCTS) simulation, the migrated model is fine-tuned based on the current data and simulation results. Taking the gradient descent-based fine-tuning method as an example, a loss function is first defined, which is used to measure the difference between the model prediction results and the actual observed data. After each simulation, the loss value is calculated based on the simulation path and the actual situation. Then, the gradient of the loss function to the model parameters is calculated through the backpropagation algorithm, and the model parameters are adjusted according to the gradient direction. For example, for parameters such as the causal weight matrix, according to the gradient descent formula Update, where is the causal weight matrix at the current moment, α is the learning rate, is the gradient of the loss function with respect to the causal weight matrix. Through multiple simulations and parameter updates, the migrated model is better adapted to the current power system security analysis problem.

[0183] In the embodiment of the present application, after completing T simulations, the score of each candidate root cause node is calculated according to the scoring formula for evaluating the candidate root cause nodes.

[0184] The node with the highest score is selected as the root cause node to complete low-latency root cause location.

[0185] Through the above steps, the designed lightweight causal inference engine combines Monte Carlo Tree Search (MCTS) with transfer learning to achieve low-latency root cause location, which can quickly and accurately identify the root cause of the problem. The scoring formula for evaluating candidate root cause nodes is as follows:

[0186]

[0187] Among them: Score(v i ): represents the candidate root cause node v i The score is used to measure the node v i The probability of being a root cause node. The higher the score, the more likely the node is to be the root cause of the problem.

[0188] v i : is a candidate root cause node. In the causal graph, there are multiple nodes that may be the root cause of the problem. These nodes are candidate root cause nodes. For example, in the power system, a generator node, a substation node, or a communication equipment node may be the root cause of the fault. They can all be candidate root cause nodes v i .

[0189] T is the number of simulations. Monte Carlo Tree Search (MCTS) is a randomized simulation-based search algorithm that explores the solution space of a problem through multiple simulations. The number of simulations determines the search accuracy and computational complexity, and typically needs to be adjusted based on the actual situation to strike a balance between computational efficiency and result accuracy.

[0190] I(v i ∈Path τ ): is an indicator function, which is defined as:

[0191]

[0192] This function is used to determine the node v i Whether it appears in the attack propagation path of the τth simulation. If it appears, the value is 1, indicating that the simulation contributes to the score of the node; otherwise, the value is 0, indicating that the simulation contributes to the score of node v i No contribution to the score.

[0193] Δt τ : is the attack propagation delay for the τth simulation. The attack propagation delay reflects the time required for the attack to propagate from the root node to other nodes. In power systems or network environments, the propagation of attacks may be affected by factors such as physical distance, communication bandwidth, and device processing capabilities. Therefore, different simulations may have different attack propagation delays.

[0194] γ is a dampening coefficient used to control the impact of attack propagation delay on node scores. A larger γ value minimizes the impact of attack propagation delay, meaning that nodes farther from the root cause node but with longer attack propagation delays can still achieve higher scores. A smaller γ value increases the impact of attack propagation delay, meaning that nodes closer to the root cause node and with shorter attack propagation delays will score higher. The value of γ is determined as follows: The dampening coefficient γ controls the impact of attack propagation delay on node scores. In practical applications, adjustments should be made based on the characteristics of the power system or network environment. First, conduct small-scale experiments with different γ values to observe their impact on root cause location. If the attack propagation delay in the power system significantly impacts root cause determination, meaning that you want to prioritize nodes closer to the root cause node and with shorter attack propagation delays, select a smaller γ value. Conversely, if the attack propagation delay has a relatively small impact, select a larger γ value. Finally, determine the optimal γ value based on analysis and experience from actual cases. For example, in multiple simulation tests, when γ=5, a relatively accurate root cause location result can be obtained while satisfying computational efficiency. In this case, it can be determined that γ=5 is an appropriate value.

[0195] In summary, the present application proposes a power system node attribution and reasoning method, which obtains the first device data of the target power system; performs a first preprocessing on the first device data, the first preprocessing including a feature extraction operation; establishes a time-varying causal graph model based on the results of the first preprocessing; and solves the time-varying causal graph model to obtain a decision strategy. This method not only improves the accuracy and efficiency of power system security monitoring, but also enables the rapid location and repair of power system faults. Specifically, through the feature extraction operation in the first preprocessing, key information related to the safety of the target power system can be effectively extracted, providing data support for the subsequent establishment of the time-varying causal graph model. At the same time, the introduction of the time-varying causal graph model enables the present application to fully consider the time factor and causal relationship in the power system, thereby more accurately reflecting the actual operating status of the power system. In addition, by solving the time-varying causal graph model, the decision strategy obtained can provide a strong guarantee for the safe operation of the power system and reduce the power outage time and economic losses caused by faults.

[0196] Example 2. In a preferred embodiment, in order to fully verify the effectiveness and superiority of the method proposed in this application, this embodiment was tested in multiple typical scenarios, covering power systems of different scales and diverse attack modes, and the experimental design details were recorded in detail.

[0197] Multi-scenario testing includes:

[0198] (1) IEEE39 node system: As a classic power system test model, it is used for basic performance testing. The system has a complex network structure and a variety of power equipment, and can simulate common power system operating conditions.

[0199] (2) Distribution network scenario: A medium-sized distribution network in actual operation is selected for testing. This distribution network contains a large number of distributed power sources, load nodes, and a complex topology. It is close to the actual application scenario and can test the performance of the method in the distribution network, such as the security attribution and reasoning capabilities under the conditions of distributed energy access and load fluctuations.

[0200] (3) Substation scenario: Testing was conducted on a large substation. The substation has numerous devices, complex communication protocols, and high electromagnetic interference. Testing in this scenario evaluated the effectiveness of the method in an environment with strong interference and complex device interactions.

[0201] The hardware configuration includes:

[0202] The test equipment uses a high-performance server equipped with an Intel Xeon Platinum 8280 processor, 64GB of memory, and an NVIDIA Tesla V100 GPU to ensure that it can efficiently process large amounts of data and complex computing tasks.

[0203] Data sources include:

[0204] (1) Sensor data: High-precision sensors are installed on the power equipment in each test scenario to collect real-time power parameters such as voltage, current, and power. The sampling frequency is 1 kHz to ensure that the real-time operating status data of the power system is obtained.

[0205] (2) SCADA system data: Obtain monitoring data from the SCADA system of the corresponding scenario, including key indicators of system operation, equipment status information, and historical operation records, etc., to provide comprehensive system operation information for security analysis.

[0206] (3) Communication data: By monitoring the communication network, we can obtain data such as interaction information and control instructions between devices transmitted based on communication protocols such as IEC61850, and analyze potential security issues in the communication process.

[0207] The comparative experiment was designed as follows:

[0208] This example uses traditional LSTM (Long Short-Term Memory) and GCN (Graph Convolutional Network) methods as comparison objects and conducts a horizontal comparison with the method of this application. To ensure fairness in the comparison, each method is tested under the same hardware environment and data source, and the key parameters of the comparison methods are reasonably set.

[0209] The comparison method parameter settings include:

[0210] (1) LSTM: The number of hidden layer units is set to 128, the learning rate is 0.001, the training batch size is 32, the Adam optimizer is used, and the number of iterations is 100. By adjusting these parameters, the LSTM achieves good performance on the test data.

[0211] (2) GCN: The convolution kernel size is 3, the number of layers is 2, the learning rate is 0.01, and the Adam optimizer is used. The number of training iterations is 80. These parameters are optimized based on the characteristics of GCN and the structure of the test data.

[0212] The test environment configuration includes:

[0213] In each test scenario, the same network environment was set, including parameters such as network bandwidth, latency, and packet loss rate. For attack patterns, the same type and intensity of attacks were simulated in each scenario, such as covert attacks (such as APT attacks), data tampering, and malicious code injection, to ensure the reliability of the comparative experimental results.

[0214] The test results and performance evaluation are as follows:

[0215] The present method and the comparative method were tested in multiple scenarios and evaluated based on multiple key performance indicators such as attribution accuracy, inference latency, and dynamic adaptability.

[0216] Attribution accuracy includes:

[0217] (1) IEEE 39-node system: The attribution accuracy of our proposed method reaches 98.2%, while the attribution accuracy of LSTM is 72.5% and that of GCN is 74.8%. Our proposed method improves the performance by 25.7% compared to LSTM and 23.4% compared to GCN.

[0218] (2) Distribution Network Scenario: The attribution accuracy of our proposed method in the distribution network scenario is 96.5%, LSTM is 68.3%, and GCN is 70.1%. This method improves by 28.2% compared to LSTM and 26.4% compared to GCN.

[0219] (3) Substation Scenario: In the substation scenario, the attribution accuracy of our proposed method is 95.8%, LSTM is 65.6%, and GCN is 67.2%. This method improves 30.2% over LSTM and 28.6% over GCN. In different scenarios, our method can capture causal relationship changes in real time through dynamic causal networks, and achieve higher attribution accuracy by combining multimodal incremental learning with multi-source heterogeneous data.

[0220] Inference latency includes:

[0221] (1) IEEE 39-node system: The inference delay of the proposed method is ≤15ms, the inference delay of LSTM is 45-60ms, and the inference delay of GCN is 30-40ms.

[0222] (2) Distribution network scenario: The inference delay of this application method is within 20ms, the LSTM delay is 50-70ms, and the GCN delay is 35-50ms.

[0223] (3) Substation Scenario: The inference delay of the proposed method in the substation scenario is approximately 25ms, while the LSTM delay is as high as 60-80ms and the GCN delay is 40-60ms. The lightweight causal inference engine proposed in this application effectively reduces the number of computational steps by optimizing the algorithm structure and data processing flow. It can meet the strict inference delay requirements of real-time control of power systems in different scenarios, while the traditional method has relatively high delay.

[0224] Dynamic adaptability includes:

[0225] (1) IEEE39 node system: This application method supports topology change frequency up to 10 Hz. When the topology change frequency reaches 5 Hz, the performance of LSTM and GCN begins to decline significantly.

[0226] (2) Distribution network scenario: In the distribution network scenario, the proposed method can adapt to the situation where the topology change frequency is 8 Hz. However, when the topology change frequency of LSTM and GCN exceeds 4 Hz, the performance indicators such as attribution accuracy and inference delay show large fluctuations.

[0227] (3) Substation scenario: For substation scenarios, the proposed method can adapt to topology change frequencies of 6 Hz. When the topology change frequency reaches 3 Hz, the performance of LSTM and GCN is difficult to meet actual requirements. The dynamic topology perception capability of the dynamic causal network proposed in this application and the adaptive characteristics of multimodal incremental learning enable it to demonstrate excellent dynamic adaptability in different scenarios, which is significantly superior to traditional methods.

[0228] This application combines dynamic causal networks with multimodal incremental learning for the first time, breaking through the spatiotemporal constraints of power scenarios. In traditional power system safety analysis methods, it is often difficult to simultaneously consider changes in temporal and spatial factors. Dynamic causal networks can capture the changes in the causal relationships between various factors in the power system over time in real time, while multimodal incremental learning can fuse data from different modalities and mine spatial features and correlation information in the data. Through this combination, this method can more comprehensively and accurately describe the operating status of the power system, providing stronger support for safety attribution and reasoning.

[0229] In response to the problem of excessive computing power requirements of traditional models, this application proposes a lightweight causal reasoning engine. Traditional causal reasoning models usually require a lot of computing resources and time to perform complex reasoning processes, which makes it difficult to meet the needs of real-time control of power systems. The lightweight causal reasoning engine reduces unnecessary calculation steps and reduces the complexity of the model by optimizing the algorithm structure and data processing flow. At the same time, combined with Monte Carlo tree search (MCTS) and transfer learning technologies, while ensuring the accuracy of reasoning, it significantly improves the reasoning efficiency and solves the problem of excessive computing power requirements of traditional models.

[0230] This method provides an efficient and explainable solution for the field of intelligent security of power systems. It can be widely used in scenarios such as smart substations and distribution network automation, and has significant technological advancement and commercial potential.

[0231] Example 3: This embodiment also provides a power system node attribution and reasoning system, including:

[0232] A data acquisition module, configured to acquire data of a first device in a target power system;

[0233] The first device data includes real-time power parameter data, monitoring data, and communication data;

[0234] A preprocessing module, configured to perform a first preprocessing on the first device data, wherein the first preprocessing includes a feature extraction operation;

[0235] The feature extraction operation is achieved by establishing a feature extraction network;

[0236] A model building module, configured to build a time-varying causal graph model based on a result of the first preprocessing;

[0237] The time-varying causal graph model is implemented by an improved first algorithm, which includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm;

[0238] The solution module is used to solve the time-varying causal graph model and obtain the decision strategy.

[0239] The above-mentioned unit modules can be embedded in or independent of the processor in the electronic device in the form of hardware, or can be stored in the memory of the electronic device in the form of software, so that the processor can call and execute the corresponding operations of the above-mentioned modules.

[0240] This embodiment also provides an electronic device, which may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 2 As shown. The electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a power system node attribution and reasoning method is implemented. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer covering the display screen, or a key, trackball or touchpad provided on the electronic device housing, or an external keyboard, touchpad or mouse.

[0241] This embodiment further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the following steps are implemented:

[0242] Acquiring data of a first device in a target power system;

[0243] The first device data includes real-time power parameter data, monitoring data, and communication data;

[0244] performing a first preprocessing on the first device data, the first preprocessing including a feature extraction operation;

[0245] The feature extraction operation is achieved by establishing a feature extraction network;

[0246] According to the result of the first preprocessing, a time-varying causal graph model is established;

[0247] The time-varying causal graph model is implemented by an improved first algorithm, which includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm;

[0248] The time-varying causal graph model is solved to obtain the decision strategy.

[0249] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, and all of these should be included in the scope of the claims of the present application.

[0250] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages.

[0251] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0252] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0253] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0254] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0255] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A power system node attribution and reasoning method, characterized in that: include: Acquiring data of a first device in a target power system; The first device data includes real-time power parameter data, monitoring data and communication data; performing a first preprocessing on the first device data, the first preprocessing comprising a feature extraction operation; The feature extraction operation is achieved by establishing a feature extraction network; Establishing a time-varying causal graph model based on the result of the first preprocessing; The time-varying causal graph model is implemented by an improved first algorithm, wherein the first algorithm includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm; The time-varying causal graph model is solved to obtain a decision strategy.

2. A power system node attribution and reasoning method according to claim 1, characterized in that: The step of establishing a time-varying causal graph model according to the result of the first preprocessing includes: Performing high-dimensional transformation on the result of the first preprocessing to obtain the result of the first preprocessing in a high-dimensional feature space; Performing an improved causal analysis algorithm on the results of the first preprocessing to obtain a plurality of causal graphs; Adjusting the plurality of causal graphs according to the improved Bayesian structure learning algorithm; A time-varying causal graph model is established based on several adjusted causal graphs.

3. A power system node attribution and reasoning method according to claim 2, characterized in that: The step of establishing a time-varying causal graph model based on the adjusted causal graphs includes: Calculating a causal weight matrix of each node according to the adjusted causal graphs; performing an estimation operation on the causal weights in the causal weight matrix; The causal weight matrix after the estimation operation is completed is used as the time-varying causal graph model.

4. A power system node attribution and reasoning method according to claim 3, characterized in that: Solving the time-varying causal graph model to obtain a decision strategy includes: Establish a scoring mechanism for evaluating candidate root cause nodes; Solving the time-varying causal graph model according to the scoring mechanism; Determine the decision-making strategy based on the solution results.

5. A power system node attribution and reasoning method according to claim 4, characterized in that: The establishment of the feature extraction network includes: performing normalization processing on the first device data of the target power system; Perform convolutional neural network feature extraction on the normalized data of the first device of the target power system; The convolutional neural network feature extraction includes the introduction of the attention mechanism; Establish a dynamic adjacency matrix and node attribute embedding results based on convolutional neural network feature extraction; Feature fusion is performed according to the dynamic adjacency matrix and the node attribute embedding result to obtain a first preprocessing result.

6. A power system node attribution and reasoning method according to claim 5, characterized in that: The dynamic adjacency matrix is a matrix showing the dynamic changes of the connection relationship between nodes in the target power system under time steps. The nodes are the devices in the target power system, and the matrix elements are the connection strength and status between the devices.

7. A power system node attribution and reasoning method and system according to claim 6, characterized in that: The node attribute embedding result is the result of mapping the node attribute information into a low-dimensional vector space; The attribute information of the node includes device type and rated power.

8. A power system node attribution and reasoning system, applying the method according to any one of claims 1 to 7, characterized in that: include: A data acquisition module, configured to acquire data of a first device in a target power system; The first device data includes real-time power parameter data, monitoring data and communication data; a preprocessing module, configured to perform a first preprocessing on the first device data, wherein the first preprocessing includes a feature extraction operation; The feature extraction operation is achieved by establishing a feature extraction network; A model building module, configured to build a time-varying causal graph model based on the result of the first preprocessing; The time-varying causal graph model is implemented by an improved first algorithm, wherein the first algorithm includes an improved causal analysis algorithm and an improved Bayesian structure learning algorithm; A solution module is used to solve the time-varying causal graph model to obtain a decision strategy.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of a power system node attribution and reasoning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a power system node attribution and reasoning method according to any one of claims 1 to 7 are implemented.

Images