Client appeal and power failure event association analysis method based on graph neural network

Abstract

The invention discloses a customer demand and power failure event association analysis method based on a graph neural network. The method comprises the following steps: collecting and preprocessing customer demand data, power failure event data, power equipment information and power grid topology data; constructing a graph structure; constructing a graph neural network model; inputting the graph structure into a graph neural network model for training; real-time customer demand data and power failure event data are collected and preprocessed, and real-time customer demand nodes and power failure event nodes are obtained; and inputting real-time customer appeal nodes and power failure event nodes into the trained graph neural network model for reasoning. The method solves the problems that an existing customer demand and power failure event correlation analysis method mainly depends on static analysis of power failure event logs and customer feedback records, but the method does not fully consider the influence of time-varying factors and geographical topology dynamic changes on customer demand and power failure event correlation; and the accuracy of an analysis result is low.

Description

Technical Field

[0001] The present invention relates to the technical field of correlation analysis between customer demands and power outage events, and specifically to a method for correlation analysis between customer demands and power outage events based on a graph neural network. Background Art

[0002] As power systems become increasingly complex and intelligent, power system operation and maintenance management faces numerous challenges and higher requirements. Customer complaints, as an important source of information reflecting user electricity experience and power system operating conditions, and their correlation with power outage events have become a key link in achieving intelligent response and precise service. Currently, existing methods for analyzing the correlation between customer complaints and power outage events mainly rely on static analysis of power outage event logs and customer feedback records. Specifically, this method collects and organizes power outage event records and customer feedback information submitted regarding power outage issues within a certain period, and uses data mining, statistical analysis, and other technical means to explore potential correlations between customer complaints and fault events. However, this method does not fully consider the impact of time-varying factors such as power load fluctuations, seasonal changes in electricity consumption, and dynamic changes in geographic topology on the correlation between customer complaints and power outage events, resulting in low accuracy and reliability of the analysis results. Summary of the Invention

[0003] In response to the above-mentioned defects, the present invention proposes a customer complaint and power outage event correlation analysis method based on graph neural network. The purpose is to solve the problem that the existing customer complaint and power outage event correlation analysis method mainly relies on static analysis of power outage event logs and customer feedback records, but this method does not fully consider the impact of time-varying factors and dynamic changes in geographic topology on the correlation between customer complaints and power outage events, resulting in low accuracy and reliability of the analysis results.

[0004] To achieve this object, the present invention adopts the following technical solutions:

[0005] A method for analyzing the correlation between customer demands and power outage events based on a graph neural network includes the following steps:

[0006] Step S1: Collect customer demand data, power outage event data, power equipment information, and grid topology data;

[0007] Step S2: Preprocess customer demand data, power outage event data, power equipment information, and grid topology data to obtain a sample data set;

[0008] Step S3: constructing a graph structure based on the common identifiers in the sample data set, wherein the graph structure includes customer demand nodes, power outage event nodes, power equipment nodes, and semantic edges;

[0009] Step S4: Build a graph neural network model;

[0010] Step S5: Input the graph structure into the graph neural network model for training to obtain a trained graph neural network model;

[0011] Step S6: collecting real-time customer demand data and power outage event data, and pre-processing the real-time customer demand data and power outage event data to obtain real-time customer demand nodes and power outage event nodes;

[0012] Step S7: Input the real-time customer demand node and the power outage event node into the trained graph neural network model for inference, and output the risk score between the real-time customer demand node and the power outage event node;

[0013] Step S8: Risk identification of power equipment is performed based on the risk scores between the real-time customer demand nodes and the power outage event nodes.

[0014] Preferably, in step S2, the customer demand data is pre-processed, specifically including the following sub-steps:

[0015] Step S21: Use the Jieba word segmentation tool to segment the customer demand data to obtain a word set after segmentation;

[0016] Step S22: using a preset stop word list to filter out common and non-informative words in the word set after segmentation;

[0017] Step S23: Use the pre-trained Word2Vec model to convert the remaining words into fixed-dimensional word vectors, and perform a mean operation on the word vectors to obtain pre-processed customer demand data;

[0018] Preprocessing of power outage event data, power equipment information, and grid topology data includes the following sub-steps:

[0019] The power outage event data, power equipment information and power grid topology data are respectively normalized to obtain normalized power outage event data, power equipment information and power grid topology data.

[0020] Preferably, in step S4, in the process of building the graph neural network model, the following steps are included:

[0021] An iterative information transmission mechanism is used to update the node features in the graph structure. The specific update formula is as follows:

[0022]

[0023] in, Represents the feature vector of node i at the kth iteration; represents the feature vector of node j at the kth iteration; W(k) represents the weight matrix at the kth iteration; b (k) Represents the bias term at the kth iteration; represents the neighborhood of node i; σ(x) represents the activation function; AGG{x} represents the aggregation function of all node features in the neighborhood of node i.

[0024] Preferably, in step S4, in the process of constructing the graph neural network model, the following steps are also included: introducing Dropout regularization and batch normalization operations in each layer of the graph neural network model.

[0025] Preferably, in step S4, during the process of constructing the graph neural network model, the following steps are also included:

[0026] Calculate the spatial edge weight between two power equipment nodes. The specific calculation formula is as follows:

[0027]

[0028] Among them, w uv represents the spatial edge weight between two power equipment nodes; d uv represents the actual distance between two power equipment nodes; τ represents the distance attenuation parameter, which is used to control the attenuation rate of the distance effect on the edge weight; exp(x) represents the exponential decay function;

[0029] Calculate the time edge weight between two customer request nodes or power outage event nodes. The specific calculation formula is as follows:

[0030]

[0031] Among them, w t represents the time edge weight between two customer demand nodes or power outage event nodes; Δt represents the time difference between two customer demand nodes or power outage event nodes; γ represents the time decay parameter, which is used to adjust the influence of the time difference on the correlation expression;

[0032] The spatial edge weight between two power equipment nodes and the temporal edge weight between two customer demand nodes or power outage event nodes are fused in a weighted summation manner to obtain the comprehensive edge weight. The specific fusion calculation formula is as follows:

[0033] w 综合 =α·w ij +(1-α)·w t ;

[0034] Among them, w 综合 represents the comprehensive edge weight; α represents the hyperparameter that adjusts the proportion of spatial and temporal influence.

[0035] Preferably, step S7 specifically includes the following sub-steps:

[0036] Step S71: Calculate the feature vector h of the real-time customer demand node using the forward propagation method of the graph neural network u And the real-time feature vector h of the power outage event node v ;

[0037] Step S72: Calculate the feature vector h of the real-time customer demand node using the cosine similarity algorithm u The feature vector h of the real-time power outage event node v The cosine similarity between them is calculated as follows:

[0038]

[0039] Among them, sim(h u ,h v ) represents the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node v The cosine similarity between

[0040] Step S73: Calculate the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv ;

[0041] Step S74: Based on the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node v The cosine similarity between the nodes and the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv , the risk score between the real-time customer demand node and the power outage event node is calculated using a weighted summation method. The specific calculation formula is as follows:

[0042]

[0043] in, represents the neighborhood of the real-time customer demand node; μ represents the hyperparameter for adjusting the similarity effect; ρ represents the hyperparameter for adjusting the comprehensive edge weight effect.

[0044] Preferably, the method further includes the following steps: using principal component analysis (PCA) to perform dimensionality reduction processing on the feature vectors of the real-time customer demand nodes and the power outage event nodes. The specific calculation formula is as follows:

[0045] ∑ u h u =λ u h u ;

[0046] ∑ v h v =λ v h v ;

[0047] Among them, ∑ u Represents the covariance matrix of real-time customer demand nodes; ∑ v Represents the covariance matrix of the real-time power outage event node; λ u Represents the eigenvalue of the covariance matrix of the real-time customer demand node; v Represents the eigenvalue of the covariance matrix of the real-time power outage event node.

[0048] Preferably, step S8 specifically includes the following sub-steps:

[0049] Step S81: setting a threshold Q of the risk score of the power equipment;

[0050] Step S82: Determine whether the risk score between the real-time customer demand node and the power outage event node is greater than or equal to Q. If so, determine that the current power equipment is a high-risk device; if not, determine that the current power equipment is a low-risk device or a risk-free device.

[0051] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0052] This solution constructs and trains a graph neural network model, then feeds real-time customer complaint nodes and power outage event nodes into the trained graph neural network for inference, enabling correlation analysis between these two nodes. Compared to existing correlation analysis methods that rely on static analysis of power outage event logs and customer feedback records, this solution leverages a graph neural network model to dynamically model these two nodes. This approach organically integrates grid topology, geographic distance, and event timing information, capturing temporal and spatial variations in the correlation between customer complaints and power outage events, thereby improving the accuracy and reliability of correlation analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flowchart of the steps of the customer demand and power outage event correlation analysis method based on graph neural network. DETAILED DESCRIPTION

[0054] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0055] A method for analyzing the correlation between customer demands and power outage events based on a graph neural network includes the following steps:

[0056] Step S1: Collect customer demand data, power outage event data, power equipment information, and grid topology data;

[0057] Step S2: Preprocess customer demand data, power outage event data, power equipment information, and grid topology data to obtain a sample data set;

[0058] Step S3: constructing a graph structure based on the common identifiers in the sample data set, wherein the graph structure includes customer demand nodes, power outage event nodes, power equipment nodes, and semantic edges;

[0059] Step S4: Build a graph neural network model;

[0060] Step S5: Input the graph structure into the graph neural network model for training to obtain a trained graph neural network model;

[0061] Step S6: collecting real-time customer demand data and power outage event data, and pre-processing the real-time customer demand data and power outage event data to obtain real-time customer demand nodes and power outage event nodes;

[0062] Step S7: Input the real-time customer demand node and the power outage event node into the trained graph neural network model for inference, and output the risk score between the real-time customer demand node and the power outage event node;

[0063] Step S8: Risk identification of power equipment is performed based on the risk scores between the real-time customer demand nodes and the power outage event nodes.

[0064] This solution uses a graph neural network-based method to analyze the correlation between customer demands and power outage events, such as Figure 1As shown, the first step is to collect customer complaint data, power outage event data, power equipment information, and grid topology data. In this embodiment, collecting this data provides a rich data foundation for the subsequent construction of the graph structure. Further explanation: customer complaint data includes the text content of user complaints, the time of complaint occurrence, the complaint channel, and sentiment; power outage event data includes the time of occurrence, fault type, affected area, and restoration time; and power equipment information includes device ID, device status, device model, installation location, and other physical attributes. Grid topology data includes the connectivity and spatial layout information between various components in the power grid, such as power plants, substations, transmission lines, distribution lines, and switchgear. The second step is to preprocess the customer complaint data, power outage event data, power equipment information, and grid topology data to obtain a sample dataset. In this embodiment, preprocessing the customer complaint data, power outage event data, power equipment information, and grid topology data helps improve the quality of these data. The third step is to construct a graph structure based on the common identifiers in the sample dataset. The graph structure includes customer complaint nodes, power outage event nodes, power equipment nodes, and semantic edges. In this embodiment, the common identifiers in the sample dataset include event IDs and device IDs. The graph structure constructed using the common identifiers in the sample dataset can closely link customer complaints, power outage events, and power equipment. Specifically, in the graph structure, customer complaint nodes carry information such as text vectors and sentiment, power outage event nodes carry information such as timestamps, regions, and fault categories, and power equipment nodes carry information such as device status, model, installation location, and other physical attributes. The connection between each node is represented by the construction of semantic edges. Semantic edges between customer complaint nodes and power outage event nodes may be assigned attributes such as time difference, event urgency, or text similarity. Semantic edges between power equipment nodes are assigned spatial distance weights based on the physical topology of the power grid and geographic distance. Edge weights can be determined using a distance decay function. The fourth step is to construct a graph neural network model. In this embodiment, the graph neural network model is a deep learning model specifically designed to process graph-structured data. By constructing a graph neural network model, we facilitate subsequent correlation analysis between customer complaint nodes and power outage event nodes. The fifth step is to input the graph structure into the graph neural network model for training, obtaining a trained graph neural network model. In this embodiment, training the graph neural network model helps improve the accuracy of subsequent correlation analysis between customer complaints and power outage events.The sixth step is to collect real-time customer complaint data and power outage event data and preprocess them to obtain real-time customer complaint nodes and power outage event nodes. In this embodiment, preprocessing the real-time customer complaint data and power outage event data, namely data vectorization, data normalization, and graph structure construction, can generate real-time customer complaint nodes and power outage event nodes, which facilitate subsequent processing by the graph neural network model. The seventh step is to input the real-time customer complaint nodes and power outage event nodes into the trained graph neural network model for inference, outputting risk scores between the real-time customer complaint nodes and power outage event nodes. In this embodiment, by inputting the real-time customer complaint nodes and power outage event nodes into the trained graph neural network model, the trained graph neural network model can deeply explore the associations between the real-time customer complaint nodes and power outage event nodes. The risk score between the real-time customer complaint nodes and power outage event nodes is obtained by weighted summation, comprehensively considering the similarity between the two nodes and the semantic edge weight. The resulting risk score can reflect the strength of the association between the real-time customer complaint nodes and power outage event nodes, facilitating subsequent risk identification of power equipment. The eighth step is to identify the risks of power equipment based on the risk scores between real-time customer demand nodes and power outage event nodes. In one embodiment, when a power outage event node, such as a transformer failure of a certain power equipment, is associated with multiple customer demand nodes, and the correlation strength, that is, the risk score, is high, such as a large number of users reporting unstable voltage, then the power equipment is a high-risk device.

[0065] This solution constructs and trains a graph neural network model, then feeds real-time customer complaint nodes and power outage event nodes into the trained graph neural network for inference, enabling correlation analysis between these two nodes. Compared to existing correlation analysis methods that rely on static analysis of power outage event logs and customer feedback records, this solution leverages a graph neural network model to dynamically model these two nodes. This approach organically integrates grid topology, geographic distance, and event timing information, capturing temporal and spatial variations in the correlation between customer complaints and power outage events, thereby improving the accuracy and reliability of correlation analysis results.

[0066] Preferably, in step S2, the customer demand data is pre-processed, specifically including the following sub-steps:

[0067] Step S21: Use the Jieba word segmentation tool to segment the customer demand data to obtain a word set after segmentation;

[0068] Step S22: using a preset stop word list to filter out common and non-informative words in the word set after segmentation;

[0069] Step S23: Use the pre-trained Word2Vec model to convert the remaining words into fixed-dimensional word vectors, and perform a mean operation on the word vectors to obtain pre-processed customer demand data;

[0070] Preprocessing of power outage event data, power equipment information, and grid topology data includes the following sub-steps:

[0071] The power outage event data, power equipment information and power grid topology data are respectively normalized to obtain normalized power outage event data, power equipment information and power grid topology data.

[0072] In this embodiment, since the customer appeal data exists in the form of unstructured text, the customer appeal data is converted into a high-dimensional vector representation by means of word segmentation, removal of stop words and word embedding to capture the semantics of the text. Further explanation, in step S21, the jieba word segmentation tool is an open source word segmentation tool commonly used in Chinese natural language processing NLP. By using the jieba word segmentation tool to perform word segmentation on the customer appeal data, efficient and accurate keyword segmentation can be achieved. In step S23, the fixed dimension is 300 dimensions. The Word2Vec model is an existing word embedding model. By using the pre-trained Word2Vec model to convert the remaining vocabulary into a fixed-dimensional word vector, it is beneficial to reduce the computational complexity when the subsequent model processes the data. Since the power outage event data, power equipment information and power grid topology data are all structured data, these data are mapped to a unified scale through normalization processing, which is beneficial to the data processing of the subsequent model.

[0073] Preferably, in step S4, in the process of constructing the graph neural network model, the following steps are included:

[0074] An iterative information transmission mechanism is used to update the node features in the graph structure. The specific update formula is as follows:

[0075]

[0076] in, Represents the feature vector of node i at the kth iteration; represents the feature vector of node j at the kth iteration; W (k) represents the weight matrix at the kth iteration; b (k) Represents the bias term at the kth iteration; represents the neighborhood of node i; σ(x) represents the activation function; AGG{x} represents the aggregation function of all node features in the neighborhood of node i.

[0077] In this embodiment, in the process of constructing the graph neural network model, the node features in the graph structure are updated by adopting an iterative information transmission mechanism, so that the node features in the graph structure can gradually integrate the information of multi-order neighbors, thereby enhancing the expressive ability of the node features in the graph structure.

[0078] Preferably, in step S4, during the construction of the graph neural network model, the following steps are further included: introducing Dropout regularization and batch normalization operations in each layer of the graph neural network model. In this embodiment, by introducing Dropout regularization and batch normalization operations in each layer of the graph neural network model, it is beneficial to suppress the overfitting phenomenon of the graph neural network model and accelerate the convergence speed of the graph neural network model.

[0079] Preferably, in step S4, during the process of constructing the graph neural network model, the following steps are also included:

[0080] Calculate the spatial edge weight between two power equipment nodes. The specific calculation formula is as follows:

[0081]

[0082] Among them, w uv represents the spatial edge weight between two power equipment nodes; d uv represents the actual distance between two power equipment nodes; τ represents the distance attenuation parameter, which is used to control the attenuation rate of the distance effect on the edge weight; exp(x) represents the exponential decay function;

[0083] Calculate the time edge weight between two customer request nodes or power outage event nodes. The specific calculation formula is as follows:

[0084]

[0085] Among them, w t represents the time edge weight between two customer demand nodes or power outage event nodes; Δt represents the time difference between two customer demand nodes or power outage event nodes; γ represents the time decay parameter, which is used to adjust the influence of the time difference on the correlation expression;

[0086] The spatial edge weight between two power equipment nodes and the temporal edge weight between two customer demand nodes or power outage event nodes are fused in a weighted summation manner to obtain the comprehensive edge weight. The specific fusion calculation formula is as follows:

[0087] w 综合 =α·w ij +(1-α)·w t ;

[0088] Among them, w 综合represents the comprehensive edge weight; α represents the hyperparameter that adjusts the proportion of spatial and temporal influence.

[0089] In this embodiment, since the power equipment node carries the geographical location information of the equipment installation, the actual distance between the two power equipment nodes can be calculated, and the spatial edge weight can be modeled using an exponential decay function. It is further explained that the distance decay parameter τ can be adjusted according to the actual geographical distribution, so that the edge weight between close-range devices is higher, while the influence of distant devices gradually weakens. Since both the customer appeal node and the power outage event node carry timestamp information, the time edge weight can be modeled by calculating the time difference between the two customer appeal nodes or power outage event nodes, and an exponential decay function can be used. It is further explained that nodes with smaller time differences have higher weights, reflecting their close connection in the time dimension. By fusing spatial edge weights and temporal edge weights, it is ensured that the graph neural network model can pay attention to the physical topological relationship between power equipment and capture the temporal dynamics of power outage events or customer appeals during the information transmission process, thereby adaptively adjusting the attention to spatiotemporal features.

[0090] Preferably, step S7 specifically includes the following sub-steps:

[0091] Step S71: Calculate the feature vector h of the real-time customer demand node using the forward propagation method of the graph neural network u And the real-time feature vector h of the power outage event node v ;

[0092] Step S72: Calculate the feature vector h of the real-time customer demand node using the cosine similarity algorithm u The feature vector h of the real-time power outage event node v The cosine similarity between them is calculated as follows:

[0093]

[0094] Among them, sim(h u ,h v ) represents the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node v The cosine similarity between

[0095] Step S73: Calculate the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv ;

[0096] Step S74: Based on the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node vThe cosine similarity between the nodes and the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv , the risk score between the real-time customer demand node and the power outage event node is calculated using a weighted summation method. The specific calculation formula is as follows:

[0097]

[0098] in, represents the neighborhood of the real-time customer demand node; μ represents the hyperparameter for adjusting the similarity effect; ρ represents the hyperparameter for adjusting the comprehensive edge weight effect.

[0099] In this embodiment, in step S71, the feature vectors of the real-time customer demand nodes and power outage event nodes are calculated by forward propagation of the graph neural network, which is conducive to subsequent similarity analysis and risk assessment. Further explanation: when the graph convolutional network (GCN) layer of the graph neural network model is used to process the real-time customer demand nodes and power outage event nodes, the real-time customer demand nodes and power outage event nodes are first formed into an input feature matrix, and then the input feature matrix is ​​input into the GCN layer of the graph neural network model. This layer uses a normalized calculation method to calculate the final feature matrix. The specific calculation method is as follows: Among them, H (k) Represents the feature matrix of all nodes in the kth iteration; represents the adjacency matrix with self-connection, express The corresponding degree matrix, whose diagonal element is the sum of the number of connections of each node; W' (k) represents the weight matrix of the GCN layer. This normalization process, by symmetric normalizing the adjacency matrix, ensures that the influence of each node is balanced during information transmission, preventing nodes with larger degrees from having an excessive impact on the aggregation results. In step S72, by quantifying the directional similarity of the feature vectors of the real-time customer demand node and the power outage event node, the semantic or attribute correlation between the customer demand and the power outage event can be accurately identified. In step S73, by calculating the comprehensive edge weights between the real-time customer demand node and the power outage event node, on the one hand, the physical topology weights of the power equipment embedded in the edge weights enhance the graph neural network model's ability to model the spatial structure of the power grid and energy flow paths; on the other hand, the temporal dynamic factors integrated into the edge weights enable the graph neural network model to learn the evolutionary laws of spatiotemporal correlations. In step S74, by calculating the risk score between the real-time customer demand node and the power outage event node using a weighted summation approach, the similarity and edge weights between the real-time customer demand node and the power outage event node are fully considered, providing a reliable reference for subsequent operation and maintenance decisions.

[0100] Preferably, the method further includes the following steps: using principal component analysis (PCA) to perform dimensionality reduction processing on the feature vectors of the real-time customer demand node and the power outage event node. The specific calculation formula is as follows:

[0101] ∑ u h u =λ u h u ;

[0102] ∑ v h v =λ v h v ;

[0103] Among them, Σ u Represents the covariance matrix of real-time customer demand nodes; Σ v Represents the covariance matrix of the real-time power outage event node; λ u Represents the eigenvalue of the covariance matrix of the real-time customer demand node; v Represents the eigenvalue of the covariance matrix of the real-time power outage event node.

[0104] Beneficial effect: In this embodiment, by using the principal component analysis (PCA) method to perform dimensionality reduction processing on the feature vectors of real-time customer demand nodes and power outage event nodes, high-dimensional data is projected into a low-dimensional space, which can reduce the data dimension and thus reduce storage requirements and computational complexity.

[0105] Preferably, step S8 specifically includes the following sub-steps: Step S81: Setting a threshold Q for the risk score of the power equipment; Step S82: Determining whether the risk score between the real-time customer demand node and the power outage event node is greater than or equal to Q. If so, the current power equipment is determined to be a high-risk device; if not, the current power equipment is determined to be a low-risk device or a risk-free device. In this embodiment, the threshold Q for the risk score of the power equipment is set to 0.8. By determining whether the risk score between the real-time customer demand node and the power outage event node is greater than or equal to Q to determine whether the current power equipment is a high-risk device, high-risk power equipment can be quickly identified.

[0106] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0107] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for analyzing the correlation between customer demands and power outage events based on graph neural networks, characterized by: The following steps are involved: Step S1: Collect customer demand data, power outage event data, power equipment information, and grid topology data; Step S2: Preprocess customer demand data, power outage event data, power equipment information, and grid topology data to obtain a sample data set; Step S3: constructing a graph structure based on the common identifiers in the sample data set, wherein the graph structure includes customer demand nodes, power outage event nodes, power equipment nodes, and semantic edges; Step S4: Build a graph neural network model; Step S5: Input the graph structure into the graph neural network model for training to obtain a trained graph neural network model; Step S6: collecting real-time customer demand data and power outage event data, and pre-processing the real-time customer demand data and power outage event data to obtain real-time customer demand nodes and power outage event nodes; Step S7: Input the real-time customer demand node and the power outage event node into the trained graph neural network model for inference, and output the risk score between the real-time customer demand node and the power outage event node; Step S8: Risk identification of power equipment is performed based on the risk scores between the real-time customer demand nodes and the power outage event nodes.

2. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 1, characterized in that: In step S2, the customer demand data is pre-processed, which specifically includes the following sub-steps: Step S21: Use the Jieba word segmentation tool to segment the customer demand data to obtain a word set after segmentation; Step S22: using a preset stop word list to filter out common and non-informative words in the word set after segmentation; Step S23: Use the pre-trained Word2Vec model to convert the remaining words into fixed-dimensional word vectors, and perform a mean operation on the word vectors to obtain pre-processed customer demand data; Preprocessing of power outage event data, power equipment information, and grid topology data includes the following sub-steps: The power outage event data, power equipment information and power grid topology data are respectively normalized to obtain normalized power outage event data, power equipment information and power grid topology data.

3. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 1, characterized in that: In step S4, the process of building the graph neural network model includes the following steps: An iterative information transmission mechanism is used to update the node features in the graph structure. The specific update formula is as follows: in, Represents the feature vector of node i at the kth iteration; represents the feature vector of node j at the kth iteration; W (k) represents the weight matrix at the kth iteration; b (k) Represents the bias term at the kth iteration; represents the neighborhood of node i; σ(x) represents the activation function; AGG{x} represents the aggregation function of all node features in the neighborhood of node i.

4. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 1, characterized in that: In step S4, in the process of building the graph neural network model, the following steps are also included: introducing Dropout regularization and batch normalization operations in each layer of the graph neural network model.

5. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 1, characterized in that: In step S4, the process of building the graph neural network model also includes the following steps: Calculate the spatial edge weight between two power equipment nodes. The specific calculation formula is as follows: Among them, w uv represents the spatial edge weight between two power equipment nodes; d uv represents the actual distance between two power equipment nodes; τ represents the distance attenuation parameter, which is used to control the attenuation rate of the distance effect on the edge weight; exp(x) represents the exponential decay function; Calculate the time edge weight between two customer request nodes or power outage event nodes. The specific calculation formula is as follows: Among them, w t represents the time edge weight between two customer demand nodes or power outage event nodes; Δt represents the time difference between two customer demand nodes or power outage event nodes; γ represents the time decay parameter, which is used to adjust the influence of the time difference on the correlation expression; The spatial edge weight between two power equipment nodes and the temporal edge weight between two customer demand nodes or power outage event nodes are fused in a weighted summation manner to obtain the comprehensive edge weight. The specific fusion calculation formula is as follows: w 综合 =α·w ij +(1-a)·w t ; Among them, w 综合 represents the comprehensive edge weight; α represents the hyperparameter that adjusts the proportion of spatial and temporal influence.

6. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 5, characterized in that: In step S7, the following sub-steps are specifically included: Step S71: Calculate the feature vector h of the real-time customer demand node using the forward propagation method of the graph neural network u And the real-time feature vector h of the power outage event node v ; Step S72: Calculate the feature vector h of the real-time customer demand node using the cosine similarity algorithm u The feature vector h of the real-time power outage event node v The cosine similarity between them is calculated as follows: Among them, sim(h u ,h v ) represents the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node v The cosine similarity between Step S73: Calculate the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv ; Step S74: Based on the feature vector h of the real-time customer demand node u The feature vector h of the real-time power outage event node v The cosine similarity between the nodes and the comprehensive edge weight w between the real-time customer demand node and the power outage event node uv , the risk score between the real-time customer demand node and the power outage event node is calculated using a weighted summation method. The specific calculation formula is as follows: in, represents the neighborhood of the real-time customer demand node; μ represents the hyperparameter for adjusting the similarity effect; ρ represents the hyperparameter for adjusting the comprehensive edge weight effect.

7. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 6, characterized in that: The following steps are also included: The principal component analysis (PCA) method is used to reduce the dimensionality of the feature vectors of the real-time customer demand node and the power outage event node. The specific calculation formula is as follows: ∑ u h u =λ u h u ; ∑ v h v =λ v h v ; Among them, Σ u Represents the covariance matrix of real-time customer demand nodes; Σ v Represents the covariance matrix of the real-time power outage event node; λ u The eigenvalue of the covariance matrix representing the real-time customer demand node; λ v Represents the eigenvalue of the covariance matrix of the real-time power outage event node.

8. The method for analyzing the correlation between customer demands and power outage events based on graph neural networks according to claim 1, characterized in that: In step S8, the following sub-steps are specifically included: Step S81: setting a threshold Q of the risk score of the power equipment; Step S82: Determine whether the risk score between the real-time customer demand node and the power outage event node is greater than or equal to Q. If so, determine that the current power equipment is a high-risk device; if not, determine that the current power equipment is a low-risk device or a risk-free device.

Images