Power grid equipment fault prediction method and system based on evolutionary structure representation learning

Abstract

The present invention provides a method and system for predicting power grid equipment faults based on evolutionary structure representation learning, which relates to the technical field of smart grid and power system fault diagnosis. The method includes collecting power grid equipment fault data, constructing a data set, and dividing the data set into a training set and a test set; constructing an evolutionary structure attention network model, the evolutionary structure attention network model includes a significant evolutionary structure sampling module, relative structure information encoding, and a node embedding encoder enhanced with structure information; using the training set to train the evolutionary structure attention network model to obtain a trained evolutionary structure attention network model; according to the characteristics of the power grid equipment fault prediction task, selecting appropriate evaluation indicators, using the test set to evaluate the trained evolutionary structure attention network model, and calculating the values ​​of various evaluation indicators. The present invention improves the accuracy and efficiency of fault prediction and provides strong technical support for the safe and stable operation of the power grid.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grid and power system fault diagnosis, and in particular to a method and system for predicting power grid equipment faults based on evolutionary structure representation learning. Background Art

[0002] With the continuous development of the power system, the stable operation of power grid equipment is of vital importance to ensure the normal operation of modern society and economic development. As the core components of the power system, the health of power grid equipment such as transformers, generators, and reactors is directly related to the stability and reliability of power supply. However, due to the complexity of power grid equipment and the tediousness of operation and maintenance, equipment failure is difficult to avoid. Once a key device fails, it may not only cause power supply interruption, but also trigger a series of chain reactions, causing serious impacts on social production and people's lives.

[0003] Traditional power grid equipment fault diagnosis methods mainly rely on the current status information of the equipment and the experience of the operation and maintenance personnel, often ignoring the complex environment in which the equipment is located and the correlation between faults. In addition, with the continuous expansion of the scale of the power grid and the improvement of the level of intelligence, the correlation between power grid equipment is becoming increasingly complex, and traditional fault diagnosis methods can no longer meet actual needs.

[0004] In recent years, the rise of dynamic graph representation learning technology has provided new ideas for power grid equipment fault diagnosis. The dynamic graph model can describe the topological relationship and fault propagation path between power grid equipment. By introducing the time dimension, it can accurately track the occurrence time of each fault, analyze the time distribution and evolution law of the fault, and thus achieve efficient and accurate diagnosis of power grid equipment faults. However, the message passing model running on a continuous dynamic network faces the challenge of dealing with the sharp increase in the number of node neighbors, which not only increases the computational complexity, but also significantly increases the amount of information processed, making it difficult to achieve real-time fault diagnosis. Summary of the invention

[0005] To this end, an embodiment of the present invention provides a power grid equipment fault prediction method and system based on evolutionary structure representation learning, which is used to solve the problem that traditional prediction methods in the prior art ignore the complex correlations between devices and the laws of fault evolution, and are difficult to cope with the computational challenges brought about by the surge in the number of node neighbors in dynamic networks.

[0006] In order to solve the above problems, an embodiment of the present invention provides a method for predicting power grid equipment faults based on evolutionary structure representation learning, the method comprising:

[0007] Step S1: Collecting power grid equipment fault data, constructing a data set according to the collected power grid equipment fault data, and dividing the data set into a training set and a test set;

[0008] Step S2: constructing an evolutionary structure attention network model, wherein the evolutionary structure attention network model includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement;

[0009] The significant evolution structure sampling module is used to construct a candidate edge set according to the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set;

[0010] The relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network;

[0011] The structure information enhanced node embedding encoder is used to generate a node representation that integrates the evolutionary structure information based on the candidate edge set sampled by the significant evolutionary structure sampling module, combining time coding and relative structure information coding;

[0012] Step S3: using the training set to train the evolutionary structure attention network model to obtain a trained evolutionary structure attention network model;

[0013] Step S4: According to the characteristics of the power grid equipment fault prediction task, select appropriate evaluation indicators, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the values ​​of various evaluation indicators.

[0014] Preferably, the step of constructing a candidate edge set according to the power grid equipment fault data specifically includes:

[0015] First, initialize the query node set, candidate edge set, and visited node set;

[0016] Then, the iterative process begins, and for a given search depth , expand the neighborhood of the query node layer by layer; in each layer, traverse each node in the current query node set, use the adjacency table search to obtain the neighbor nodes and edges of these nodes before the query time, and ensure that the searched nodes are not visited repeatedly; the newly discovered neighbor nodes and edges are added to the visited node set and the candidate edge set respectively; as the iteration proceeds, the query node set continues to expand until the predetermined search depth is reached ;

[0017] Finally, the output contains the query nodes before the query time The set of candidate edges for all edges in the hop neighborhood.

[0018] Preferably, the constructed candidate edge set samples significant edges according to the change of von Neumann entropy caused by the edges to obtain the sampled candidate edge set, which specifically includes:

[0019] First, the constructed candidate edge set is sorted by time and divided into multiple windows with equal time intervals based on a given time domain; in each window, the contribution of each edge to the change of von Neumann entropy is calculated;

[0020] Subsequently, the edges are sorted according to the calculated entropy change, so as to give priority to the edges that have a greater impact on the network structure. During the sampling process, a sampling ratio and a maximum number of neighbors for each node are set to ensure the effectiveness of the sampling and the balance of the network structure. Specifically, edges are selected in order from the sorted edge list. Only when the current number of neighbors of both endpoints of the edge has not reached the maximum value, the edge will be added to the sampling set. This process continues until the sampling number limit for each time window is reached or all eligible edges are traversed.

[0021] Finally, the sampling results of all time windows are summarized and the final set of sampled candidate edges is output.

[0022] Preferably, the method for calculating the contribution of each edge to the change in von Neumann entropy is:

[0023] ;

[0024] In the formula, represents the contribution of each edge to the change in von Neumann entropy, , They represent the continuous dynamic graph network in time and The structure of time, represents an edge; , Respectively represent nodes , The degree of change; Represents a collection of nodes.

[0025] Preferably, the candidate edge set sampled by the significant evolutionary structure sampling module is combined with time coding and relative structure information coding to generate a node representation integrating the evolutionary structure information, specifically including:

[0026] First, the evolutionary structured attention network model is Layer, through time Edge , from the node arrive Generate Message for:

[0027] ;

[0028] In the formula, || is the connection operation, is the query timestamp, is the random Fourier time code, It is the edge The relative structural information encoding of It is The weight metric used for message passing on the layer, Is a node In the The hidden representation output on the layer, ,in is the dimension of node embedding;

[0029] Then, the Evolved Structure Attention Network model aggregates all messages using a multi-head attention mechanism and adds temporal information to generate nodes The hidden embedding representation of :

[0030] ;

[0031] in

[0032] ;

[0033] ;

[0034] ;

[0035] In the formula, MultiHeadAttn represents the multi-head attention mechanism, is the maximum aggregate number of neighbors in a 1-hop neighborhood, , and It is a learnable parameter matrix, which is used to aggregate the neighborhood messages of this layer to form the Key, Query, and Value matrices required by the attention mechanism. Represent the Key, Query, and Value matrices respectively, The code representing the current query time;

[0036] Finally, the evolutionary structure attention network model calculates the similarity score between the query and all the keys, obtains the output vector of each head, concatenates the output of each head, and obtains the final node hidden embedding of this layer through the final linear transformation. .

[0037] Preferably, the evolutionary structure attention network model uses a score function in the decoding stage. To predict at the query timestamp Time Node and The probability of a link between:

[0038] ;

[0039] Among them, MLP represents multi-layer perceptron, Represents the activation function.

[0040] Preferably, the training set is used to train the evolutionary structure attention network model. In order to effectively train the node embedding encoder with enhanced structure information, a training loss function is designed, which is expressed as follows:

[0041] ;

[0042] in, represents the training loss function, Indicates that at a given time The two nodes observed below The connection between represents the candidate edge set, represents a negative sample node, represents the probability distribution, represents the number of negative samples, represents mathematical expectation.

[0043] The embodiment of the present invention further provides a power grid equipment fault prediction system based on evolutionary structure representation learning, which is used to implement the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning, and specifically includes:

[0044] The data collection and division module is used to collect power grid equipment fault data, construct a data set based on the collected power grid equipment fault data, and divide the data set into a training set and a test set;

[0045] A model building module, used to build an evolutionary structure attention network model, wherein the evolutionary structure attention network model includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement;

[0046] The significant evolution structure sampling module is used to construct a candidate edge set according to the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set;

[0047] The relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network;

[0048] The structure information enhanced node embedding encoder is used to generate a node representation that integrates the evolutionary structure information based on the candidate edge set sampled by the significant evolutionary structure sampling module, combining time coding and relative structure information coding;

[0049] A model training module is used to train the evolutionary structure attention network model using the training set to obtain a trained evolutionary structure attention network model;

[0050] The model evaluation module is used to select appropriate evaluation indicators according to the characteristics of the power grid equipment fault prediction task, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the values ​​of various evaluation indicators.

[0051] An embodiment of the present invention also provides an electronic device, which includes a processor, a memory and a bus system, wherein the processor and the memory are connected through the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning.

[0052] An embodiment of the present invention also provides a computer storage medium, which stores a computer software product. The computer software product includes several instructions for enabling a computer device to execute the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning.

[0053] It can be seen from the above technical solutions that the present invention has the following beneficial effects:

[0054] (1) Efficiently capture dynamic evolution features: By introducing a significant evolution structure sampling module based on von Neumann entropy, the present invention can efficiently identify key structural changes in power grid equipment fault data and prioritize edges with high significance. This method not only reduces computational complexity, but also significantly improves the model's ability to capture dynamic network evolution features, making fault prediction more accurate.

[0055] (2) Enhanced structural information representation: By using relative structural information encoding technology, the present invention introduces more details and in-depth information about network structure changes during message transmission. This enhanced structural information representation method enables the model to have a deeper understanding of the dynamic topological characteristics of power grid equipment fault data, thereby improving the accuracy and robustness of fault prediction.

[0056] (3) Wide applicability and high efficiency: Through experimental verification on different data sets, the present invention has demonstrated its versatility and high efficiency in long-term evolution networks. Compared with multiple advanced baseline models, the present invention has shown significant advantages in power grid equipment fault prediction tasks and has broad application prospects and high practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the implementation cases of the present invention or the technical solutions in the prior art, the following is a brief description of the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0058] Figure 1 A flowchart of a power grid equipment fault prediction method based on evolutionary structure representation learning provided in an embodiment;

[0059] Figure 2 A block diagram of a power grid equipment fault prediction system based on evolutionary structure representation learning provided in an embodiment. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Embodiment 1

[0061] In order to solve the problem of power system, traditional prediction methods ignore the complex relationship between devices and the evolution law of faults, and are difficult to cope with the computational challenges brought by the surge in the number of node neighbors in dynamic networks. Figure 1 As shown, an embodiment of the present invention proposes a method for predicting power grid equipment faults based on evolutionary structure representation learning, the method comprising:

[0062] Step S1: Collecting power grid equipment fault data, constructing a data set according to the collected power grid equipment fault data, and dividing the data set into a training set and a test set;

[0063] Step S2: constructing an evolutionary structure attention network model, which includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement;

[0064] The significant evolution structure sampling module is used to construct a candidate edge set based on the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set;

[0065] Relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network;

[0066] The node embedding encoder with enhanced structural information is used to generate node representations that integrate the evolving structural information based on the candidate edge set sampled by the significant evolving structural sampling module, combining temporal encoding with relative structural information encoding;

[0067] Step S3: using the training set to train the evolutionary structure attention network model to obtain a trained evolutionary structure attention network model;

[0068] Step S4: According to the characteristics of the power grid equipment fault prediction task, select appropriate evaluation indicators, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the values ​​of various evaluation indicators.

[0069] From the above technical solution, it can be seen that the present invention provides a method for predicting power grid equipment faults based on evolutionary structure representation learning. First, a data set containing power grid equipment fault information is constructed, and an evolutionary structure attention network model is designed, in which the significant evolutionary structure sampling module accurately identifies and prioritizes edges with significant structural changes, effectively reducing the amount of calculation and improving the pertinence of data processing; the relative structure information encoding deeply explores the relative differences in network structure changes, enhancing the model's ability to capture dynamic topological characteristics; finally, the node embedding encoder enhanced by structure information fuses spatiotemporal evolution information to generate node representations rich in fault precursors. This series of steps significantly improves the accuracy and efficiency of fault prediction, and provides strong technical support for the safe and stable operation of the power grid.

[0070] In step S1, power grid equipment fault data is collected, a data set is constructed according to the collected power grid equipment fault data, and the data set is divided into a training set and a test set.

[0071] In this embodiment, in order to predict and analyze the faults of power grid equipment, the present invention maps the collected power grid equipment fault data into a graph structure. The following are the specific steps for constructing a temporal graph based on the power grid equipment fault data:

[0072] Data collection and preprocessing: Collect fault data of power grid equipment, including fault time, fault type, impact range, connection relationship between equipment, etc. Clean and preprocess the data to remove noise and outliers to ensure data accuracy and consistency.

[0073] Define nodes: Create corresponding nodes in the graph based on the unique identifier or type of the grid equipment. Assign necessary attributes to each node, such as equipment type, geographic location, etc.

[0074] Define edges: Create edges in the graph based on the connection relationship between devices or potential fault propagation paths. The direction of the edge can be determined based on the actual situation of fault propagation. Assign necessary attributes to each edge, such as fault propagation probability, delay time, etc.

[0075] Consider the time factor: Introduce the concept of timestamp or time window in the graph to indicate the time when the fault occurs or the predicted time range. The time factor can be used as one of the attributes of the node or edge, or a special time node can be created in the graph to represent different time points.

[0076] Constructing a temporal graph: Considering the impact of time on fault propagation, a temporal graph is constructed. In a temporal graph, the edges not only represent the connection relationship between devices, but also contain timestamp information, indicating the time of fault propagation. The temporal graph can more accurately simulate the dynamic propagation process of power grid equipment faults.

[0077] In step S2, an Evolutionary Structure Attention Network (ESAN) model is constructed. The ESAN model includes a saliency evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement.

[0078] In this embodiment, the significant evolution structure sampling module is used to construct a candidate edge set based on the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set. The significant evolution structure sampling module focuses on identifying and sampling the most influential edges in the network structure. The importance of the edge is determined by measuring the von Neumann entropy changes caused by the edge, thereby simplifying the network structure while maintaining the evolutionary dynamics of the network.

[0079] Specifically, the present invention designs a candidate edge set search algorithm. The algorithm first initializes the query node set, the candidate edge set and the visited node set. Then it enters an iterative process, for a given search depth , expand the neighborhood of the query node layer by layer; in each layer, traverse each node in the current query node set, use the adjacency table search to obtain the neighbor nodes and edges of these nodes before the query time, and ensure that the searched nodes are not visited repeatedly; the newly discovered neighbor nodes and edges are added to the visited node set and the candidate edge set respectively; as the iteration proceeds, the query node set continues to expand until the predetermined search depth is reached The final output contains the query nodes before the query time The set of candidate edges for all edges in the hop neighborhood.

[0080] After the candidate edge set is constructed, the present invention designs a significant evolution structure sampling algorithm. The algorithm first sorts the constructed candidate edge set by time and divides it into multiple windows with equal time intervals based on a given time domain. In each window, the contribution of each edge to the change in von Neumann entropy is calculated. This metric reflects the importance of the edge in the dynamic change of the network structure. Then the edges are sorted according to the calculated entropy change so as to give priority to the edges that have a greater impact on the network structure. During the sampling process, a sampling ratio and a maximum number of neighbors for each node are set to ensure the effectiveness of the sampling and the balance of the network structure. Specifically, edges are selected in order from the sorted edge list, and only when the current number of neighbors of both endpoints of the edge does not reach the maximum value, the edge will be added to the sampling set. This process continues until the sampling number limit of each time window is reached or all qualified edges are traversed. Finally, the sampling results of all time windows are summarized and the final sampled candidate edge set is output.

[0081] The method for calculating the contribution of each edge to the change in von Neumann entropy is:

[0082] ;

[0083] In the formula, represents the contribution of each edge to the change in von Neumann entropy, , They represent the continuous dynamic graph network in time and The structure of time, represents an edge; , Respectively represent nodes , The degree of change; Represents a collection of nodes.

[0084] Furthermore, relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network. Relative structural information encoding is implemented through a three-dimensional multi-hot encoding vector, where each element of the three-dimensional encoding vector depends on whether the following key constraints are met, and is assigned a value of 1 if satisfied, and a value of 0 if not satisfied:

[0085] (1) Whether the interaction frequency between nodes on the edge is at the top of the candidate edge set ?

[0086] (2) Is the von Neumann entropy change associated with the edge in the front of the candidate edge set? ?

[0087] (3) Is the timestamp of the edge the closest to the query timestamp in the edge set? ?

[0088] in, Represents an adjustable percentage relative threshold, allowing for flexible and dynamic analysis of network structure. By encoding this relative structural difference, the ESAN model can distinguish between these different types of connections, thereby gaining a more nuanced understanding of the network architecture. Relative structural information encoding leverages the concept of relative structural differences between edges, focusing on the differences in the structural roles of edges in the network, providing a comprehensive and multidimensional view of network interrelationships and dynamics.

[0089] Furthermore, the node embedding encoder with enhanced structural information is used to generate node representations that incorporate the evolutionary structural information based on the candidate edge set sampled by the saliency evolutionary structure sampling module, combining temporal encoding with relative structural information encoding. First, the evolutionary structure attention network model is used in the first Layer, through time Edge , from the node arrive Generate Message for:

[0090] ;

[0091] In the formula, || is the connection operation, is the query timestamp, is the random Fourier time code, It is the edge The relative structural information encoding of It is The weight metric used for message passing on the layer, Is a node In the The hidden representation output on the layer, ,in is the dimension of node embedding;

[0092] Then, the Evolved Structure Attention Network model aggregates all messages using a multi-head attention mechanism and adds temporal information to generate nodes The hidden embedding representation of :

[0093] ;

[0094] in

[0095] ;

[0096] ;

[0097] ;

[0098] In the formula, MultiHeadAttn represents the multi-head attention mechanism, is the maximum aggregate number of neighbors in a 1-hop neighborhood, , and It is a learnable parameter matrix, which is used to aggregate the neighborhood messages of this layer to form the Key, Query, and Value matrices required by the attention mechanism. Represent the Key, Query, and Value matrices respectively, The code representing the current query time;

[0099] Finally, the evolutionary structure attention network model calculates the similarity score between the query and all the keys, obtains the output vector of each head, concatenates the output of each head, and obtains the final node hidden embedding of this layer through the final linear transformation. .

[0100] Furthermore, in the decoding stage, the score function To predict at the query timestamp Time Node and The probability of a link between:

[0101] ;

[0102] Among them, MLP represents multi-layer perceptron, Represents the activation function.

[0103] At the same time, in order to effectively train the node embedding encoder with enhanced structural information, the present invention incorporates the idea of ​​negative sampling and designs a training loss function, which is expressed as follows:

[0104] ;

[0105] in, represents the training loss function, Indicates that at a given time The two nodes observed below The connection between represents the candidate edge set, represents a negative sample node, represents the probability distribution, represents the mathematical expectation, Represents the number of negative samples, which are distributed from the node space The introduction of negative sampling helps the model distinguish between real edges and randomly generated false edges, thereby improving its identification ability. Through such a loss function design, the model can gradually learn the structural characteristics and evolution laws on the dynamic network during the training process, providing a solid performance foundation for the model.

[0106] In step S3, the training set is used to train the evolutionary structure attention network model to obtain a trained evolutionary structure attention network model.

[0107] In step S4, according to the characteristics of the power grid equipment fault prediction task, appropriate evaluation indicators are selected, the trained evolutionary structure attention network model is evaluated using the test set, and the values ​​of various evaluation indicators are calculated.

[0108] In order to further illustrate the advantages of the present invention, a specific case is described below.

[0109] 1. Introduction to experimental dataset

[0110] This paper uses five highly representative and rich datasets, covering different evolution time ranges and network structure sizes, to demonstrate the significant contribution of this paper - the effectiveness of alleviating the neighborhood explosion problem in dynamic network representation learning. The five datasets are: MathOverflow dataset, BitcoinOTC dataset, BitcoinAlpha dataset, Wikipedia dataset, and Reddit dataset.

[0111] As shown in Table 1, the network statistics show the diversity of the evolution time range and network structure scale. In particular, this paper uses a very large dense dataset Reddit to examine the ability of the ESAN model to handle large-scale, high-density dynamic network data. Through the detailed analysis and application of these datasets, this paper aims to fully demonstrate the unique advantages of the ESAN model in capturing and understanding dynamic network changes, and its wide adaptability and efficiency in solving practical problems, and demonstrate the excellent performance and practical value of the ESAN model in handling continuous dynamic network representation learning tasks.

[0112] Table 1 Dataset statistics

[0113]

[0114] 2. Introduction to Benchmark Method

[0115] In order to fairly evaluate the performance of the ESAN model in processing dynamic network representation learning tasks, this paper introduces six mainstream benchmark methods for comparison. The benchmark models are introduced in detail as follows:

[0116] (1) The JODIE method estimates the future node embedding trajectory by updating the latent states of two nodes in the interaction through two interacting recurrent neural networks.

[0117] (2) The DyRep method integrates the sequence model with the attention message passer and exploits the two-hop neighborhood information to learn time-aware node embeddings.

[0118] (3) The TGAT method samples and aggregates multi-hop neighborhood information to learn time-aware node embeddings. In this method, the temporal information is preserved through random Fourier time coding when information is passed.

[0119] (4) The TGN method proposes a general framework based on message passing to learn CTDGs, which has a node memory update mechanism and inherits and combines the key designs of JODIE and TGAT.

[0120] (5) The CAWs method uses recurrent neural networks and pooling modules to encode and aggregate anonymous space-time walks to learn interactive representations.

[0121] (6) The TDLG method constructs an edge graph of the network, builds an adjacency matrix based on the edge graph, and performs link prediction by training a logistic regression classifier.

[0122] 3. Experimental evaluation

[0123] (1) Direct link prediction task

[0124] The present invention uses the transductive link prediction (Transductive LinkPrediction) in the link prediction task as the model performance evaluation task. Transductive link prediction focuses on the generalization from a specific training sample to a specific test sample, that is, during the test process, invisible nodes in the training phase will not be introduced, and future unknown links will be predicted between existing nodes. This setting assumes that the network will not add unseen nodes in the future test period to accurately evaluate the model's ability to capture the historical evolution information of the network, and fairly quantify the model's ability to capture evolution laws within a limited neighborhood size.

[0125] (2) Evaluation indicators

[0126] The present invention uses common AP and AUC indicators to evaluate the prediction effect.

[0127] 4. Experimental parameter settings

[0128] Regarding the hyperparameter settings of the ESAN model, the main hyperparameter settings are as follows: (1) Attention head = 2, (2) Number of stacked layers of graph encoding network , (3) Node hidden layer embedding dimension , (4) Sampling search depth = 2 (5) Constraint on the maximum number of sampling neighborhoods = 80. At the same time, the sampling ratio Frequency, entropy change, and time relative thresholds for encoding relative structural information , The different settings on each dataset are shown in Table 2.

[0129] Table 2 ESAN model hyperparameter settings Unit: %

[0130]

[0131] In terms of experimental environment parameter settings, all model experiments were conducted in the Tesla A100 GPU and Intel 8358 Xeon CPU experimental environments, using the Adam optimizer for training, with the initial learning rate set to 0.001, 50 training rounds, and an early stopping strategy for training. In order to increase the credibility and stability of the experimental results, the mean and standard deviation in the experimental results come from 5 repeated experiments.

[0132] 5. Performance Analysis

[0133] Table 3 lists the AUC index results of the model in the direct push link prediction task of five data sets, and Table 4 shows the corresponding AP index results. In these tables, the present invention uses bold fonts to mark the effect of the optimal model under each data set, underlines the effect of the suboptimal model, and the horizontal line "——" indicates that the model cannot be extended to the data set in the experimental environment. The results are from the mean and variance of 5 experiments.

[0134] Table 3 AUC results of direct link prediction Unit: %

[0135]

[0136] Table 4 Direct-push link prediction AP results Unit: %

[0137]

[0138] The results in Tables 3 and 4 together show that the ESAN model consistently outperforms or is comparable to the performance of competing models in all cases. The ESAN model shows robust performance in both networks with long and short evolution times, in stark contrast to the baseline model, which has significant differences in performance under these network conditions, with many networks performing poorly on long-term evolving datasets, suffering from the expansive nature of network evolution. Specifically for each dataset, the AUC and AP indicators of the ESAN model on the MathOverflow dataset reached 98.03%±0.32 and 98.35%±0.39, respectively, indicating that the ESAN model is particularly advantageous when dealing with networks with long-term evolution characteristics. On other datasets, the ESAN model also shows performance that is superior to or at least comparable to the best competing models: the AP indicators of the ESAN model on the BitcoinOTC and BitcoinAlpha datasets reached 98.30%±0.38 and 97.55±0.65, respectively, which are significantly higher than other models. On the Wikipedia and Reddit datasets, the ESAN model is the second-best model, and the AP and AUC indicators are within 0.5% of the best models CAW and TGN in the field. Compared with the huge variance of CAW and TGN in each dataset, the ESAN model performs more stably on each dataset. In particular, the cutting-edge model TDLG published at the 2023 ICLR conference is limited by the need to construct an edge graph adjacency matrix with edges as nodes, which makes it difficult to run on the large-scale intensive dataset Reddit, while the ESAN model can be well expanded and achieve good results, which further confirms the versatility and efficiency of the ESAN model in long-term evolving networks. In addition, the ESAN model significantly outperforms several other benchmark models, such as JODIE, DyRep, TGAT, TGN, and CAW, in all datasets, further verifying its superiority in direct-push link prediction tasks. Embodiment 2

[0139] like Figure 2 As shown, the present invention provides a power grid equipment fault prediction system based on evolutionary structure representation learning, which is used to implement the power grid equipment fault prediction method based on evolutionary structure representation learning of the above-mentioned embodiment 1, and specifically includes:

[0140] The data collection and division module 100 is used to collect power grid equipment fault data, construct a data set according to the collected power grid equipment fault data, and divide the data set into a training set and a test set;

[0141] A model building module 200 is used to build an evolutionary structure attention network model, which includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement;

[0142] The significant evolution structure sampling module is used to construct a candidate edge set based on the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set;

[0143] Relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network;

[0144] The node embedding encoder with enhanced structural information is used to generate node representations that integrate the evolving structural information based on the candidate edge set sampled by the significant evolving structural sampling module, combining temporal encoding with relative structural information encoding;

[0145] A model training module 300 is used to train the evolutionary structure attention network model using a training set to obtain a trained evolutionary structure attention network model;

[0146] The model evaluation module 400 is used to select appropriate evaluation indicators according to the characteristics of the power grid equipment fault prediction task, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the value of each evaluation indicator.

[0147] A power grid equipment fault prediction system based on evolutionary structure representation learning in this embodiment is used to implement the aforementioned power grid equipment fault prediction method based on evolutionary structure representation learning. Therefore, the specific implementation method of the power grid equipment fault prediction system based on evolutionary structure representation learning can be seen in the implementation method part of the power grid equipment fault prediction method based on evolutionary structure representation learning in the previous text. For example, the data acquisition and division module 100, the model construction module 200, the model training module 300, and the model evaluation module 400 are respectively used to implement steps S1, S2, S3, and S4 in the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part. In order to avoid redundancy, it will not be repeated here. Embodiment 3

[0148] An embodiment of the present invention provides an electronic device, which includes a processor, a memory and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning. Embodiment 4

[0149] An embodiment of the present invention provides a computer storage medium storing a computer software product. The computer software product includes several instructions for enabling a computer device to execute the above-mentioned power grid equipment fault prediction method based on evolutionary structure representation learning.

[0150] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may 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.) containing computer-usable program codes.

[0151] 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 generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. 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.

[0152] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate 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 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0153] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from them are still within the protection scope of the invention.

Claims

1. A method for predicting power grid equipment faults based on evolutionary structure representation learning, characterized in that: include: Step S1: Collect power grid equipment fault data, build a data set based on the collected power grid equipment fault data, and divide the data set into a training set and a test set, specifically including: Mapping the collected power grid equipment fault data to the graph structure, and constructing a temporal graph based on the power grid equipment fault data includes: Collect fault data of power grid equipment, including fault time, fault type, impact range, and connection relationship between equipment; According to the unique identifier or type of the grid equipment, the corresponding nodes are created in the graph, and necessary attributes are assigned to each node, including equipment type and geographical location; Create edges in the graph based on the connection relationship between devices or potential fault propagation paths. The direction of the edge is determined based on the actual situation of fault propagation. Necessary attributes are assigned to each edge, including fault propagation probability and delay time. Step S2: constructing an evolutionary structure attention network model, wherein the evolutionary structure attention network model includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement; The significant evolution structure sampling module is used to construct a candidate edge set according to the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set; The relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network; The structure information enhanced node embedding encoder is used to generate a node representation that integrates the evolutionary structure information based on the candidate edge set sampled by the significant evolutionary structure sampling module, combining time coding and relative structure information coding; Step S3: using the training set to train the evolutionary structure attention network model to obtain a trained evolutionary structure attention network model; Step S4: According to the characteristics of the power grid equipment fault prediction task, select appropriate evaluation indicators, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the values ​​of various evaluation indicators.

2. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 1 is characterized in that: The step of constructing a candidate edge set according to the power grid equipment fault data specifically includes: First, initialize the query node set, candidate edge set, and visited node set; Then, the iterative process begins, and for a given search depth , expand the neighborhood of the query node layer by layer; in each layer, traverse each node in the current query node set, use the adjacency table search to obtain the neighbor nodes and edges of these nodes before the query time, and ensure that the searched nodes are not visited repeatedly; the newly discovered neighbor nodes and edges are added to the visited node set and the candidate edge set respectively; as the iteration proceeds, the query node set continues to expand until the predetermined search depth is reached ; Finally, the output contains the query nodes before the query time The set of candidate edges for all edges in the hop neighborhood.

3. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 1 is characterized in that: Based on the constructed candidate edge set, sampling significant edges according to the von Neumann entropy change caused by the edges to obtain the sampled candidate edge set specifically includes: First, the constructed candidate edge set is sorted by time and divided into multiple windows with equal time intervals based on a given time domain; in each window, the contribution of each edge to the change of von Neumann entropy is calculated; Subsequently, the edges are sorted according to the calculated entropy change, so as to give priority to the edges that have a greater impact on the network structure. During the sampling process, a sampling ratio and a maximum number of neighbors for each node are set to ensure the effectiveness of the sampling and the balance of the network structure. Specifically, edges are selected in order from the sorted edge list. Only when the current number of neighbors of both endpoints of the edge has not reached the maximum value, the edge will be added to the sampling set. This process continues until the sampling number limit for each time window is reached or all eligible edges are traversed. Finally, the sampling results of all time windows are summarized and the final set of sampled candidate edges is output.

4. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 3 is characterized in that: The method for calculating the contribution of each edge to the change in von Neumann entropy is: ； In the formula, represents the contribution of each edge to the change in von Neumann entropy, , They represent the continuous dynamic graph network in time and The structure of time, represents an edge; , Respectively represent nodes , The degree of change; Represents a collection of nodes.

5. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 1, characterized in that: The candidate edge set sampled by the significant evolutionary structure sampling module is combined with time coding and relative structure information coding to generate a node representation integrating evolutionary structure information, specifically including: First, the evolutionary structured attention network model is Layer, through time Edge , from the node arrive Generate Message for: ； In the formula, || is the connection operation, is the query timestamp, is the random Fourier time code, It is the edge The relative structural information encoding of It is The weight metric used for message passing on the layer, Is a node In the The hidden representation output on the layer, ,in is the dimension of node embedding; Then, the Evolved Structure Attention Network model aggregates all messages using a multi-head attention mechanism and adds temporal information to generate nodes The hidden embedding representation of : ； in ； ； ； In the formula, MultiHeadAttn represents the multi-head attention mechanism, is the maximum aggregate number of neighbors in a 1-hop neighborhood, , and It is a learnable parameter matrix, which is used to aggregate the neighborhood messages of this layer to form the Key, Query, and Value matrices required by the attention mechanism. Represent the Key, Query, and Value matrices respectively, The code representing the current query time; Finally, the evolutionary structure attention network model calculates the similarity score between the query and all the keys, obtains the output vector of each head, concatenates the output of each head, and obtains the final node hidden embedding of this layer through the final linear transformation. .

6. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 5 is characterized in that: The evolutionary structure attention network model uses the score function in the decoding stage. To predict at the query timestamp Time Node and The probability of a link between: ； Among them, MLP represents multi-layer perceptron, Represents the activation function.

7. The power grid equipment fault prediction method based on evolutionary structure representation learning according to claim 1 is characterized in that: The training set is used to train the evolutionary structure attention network model. In order to effectively train the node embedding encoder with enhanced structure information, a training loss function is designed, which is expressed as follows: ； in, represents the training loss function, Indicates that at a given time The two nodes observed below The connection between represents the candidate edge set, represents a negative sample node, represents the probability distribution, represents the number of negative samples, represents mathematical expectation.

8. A power grid equipment fault prediction system based on evolutionary structure representation learning, characterized in that: The system is used to implement the power grid equipment fault prediction method based on evolutionary structure representation learning as described in any one of claims 1 to 7, specifically comprising: The data collection and division module is used to collect power grid equipment fault data, construct a data set based on the collected power grid equipment fault data, and divide the data set into a training set and a test set; A model building module, used to build an evolutionary structure attention network model, wherein the evolutionary structure attention network model includes a significant evolutionary structure sampling module, a relative structure information encoding, and a node embedding encoder for structure information enhancement; The significant evolution structure sampling module is used to construct a candidate edge set according to the power grid equipment fault data, and based on the constructed candidate edge set, sample significant edges according to the von Neumann entropy changes caused by the edges to obtain the sampled candidate edge set; The relative structural information encoding is used to analyze the relative differences in network structure changes, introduce more details and depth to the message passing representation, and enhance the model's ability to understand and capture the dynamic topological characteristics of the network; The structure information enhanced node embedding encoder is used to generate a node representation that integrates the evolutionary structure information based on the candidate edge set sampled by the significant evolutionary structure sampling module, combining time coding and relative structure information coding; A model training module is used to train the evolutionary structure attention network model using the training set to obtain a trained evolutionary structure attention network model; The model evaluation module is used to select appropriate evaluation indicators according to the characteristics of the power grid equipment fault prediction task, use the test set to evaluate the trained evolutionary structure attention network model, and calculate the values ​​of various evaluation indicators.

9. An electronic device, characterized in that: The electronic device includes a processor, a memory and a bus system, the processor and the memory are connected through the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the power grid equipment fault prediction method based on evolutionary structure representation learning as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: The computer storage medium stores a computer software product, and the computer software product includes a number of instructions for enabling a computer device to execute the power grid equipment fault prediction method based on evolutionary structure representation learning as described in any one of claims 1 to 7.

Images