Fault prediction and operation and maintenance method of subway power distribution system based on big data analysis

Abstract

The present invention discloses a method for predicting and maintaining subway power distribution system faults based on big data analysis. The method comprises the following steps: S1: data preprocessing; S2: constructing a distribution network topology map; S3: extracting node spatial embeddings based on a multi-head GAT; S4: generating temporal embeddings using a bidirectional LSTM and temporal attention; S5: fusing spatiotemporal features using cross-spatiotemporal attention; S6: training an autoencoder to learn a reconstruction function using only normal samples; S7: adaptively generating a threshold within a sliding window using a GMM and mapping the reconstruction error to a risk score; and S8: automatically dividing suspected fault areas according to connected components when the risk score exceeds a dynamic threshold, and issuing warning information containing the risk level, suspected fault area, and recommended actions. This method can achieve forward-looking prediction of subway power distribution system fault risks, enabling operators to take preventive measures in advance, thereby ensuring the efficient and stable operation of the subway power distribution system.

Description

Technical Field

[0001] The present invention relates to the field of power grid fault monitoring and operation and maintenance technology, and in particular to a subway power distribution system fault prediction and operation and maintenance method based on big data analysis. Background Art

[0002] As the backbone of urban public transportation, the safe, stable, and efficient operation of the subway is crucial to the smooth functioning of the city. The power distribution system is the lifeline of subway operations, responsible for providing reliable power to key equipment such as train traction, signaling, communications, lighting, and environmental control. Failures in the power distribution system can lead to train delays and interruptions to station services, or even worse, to safety incidents, resulting in significant economic losses and social impact. To ensure the safe and stable operation of the subway network, proactive prediction of the risks of subway power distribution system failures is crucial. Summary of the Invention

[0003] Based on the above, the purpose of the present invention is to propose a subway power distribution system fault prediction and operation and maintenance method based on big data analysis, which can realize the forward-looking prediction of the failure risk of the subway power distribution system, so as to take preventive measures in advance and ensure the efficient and stable operation of the subway power distribution system.

[0004] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0005] The present invention provides a subway power distribution system fault prediction and operation and maintenance method based on big data analysis, comprising:

[0006] S1: Obtain historical and real-time monitoring data of the distribution network and preprocess the acquired data samples;

[0007] S2: Convert the distribution network topology into a topology graph data model. Map each power device in the distribution network as a node of the graph based on the physical structure of the distribution network, with feeders as edges of the graph, and assign pre-processed data to corresponding node attributes.

[0008] S3: Based on the topological graph, a multi-head graph attention network is used to aggregate features of the neighborhood of each node to obtain the spatial embedding vector of each node;

[0009] S4: Input the monitoring time series data of each node into the long short-term memory network, extract the time context features, and perform time step aggregation to obtain the time embedding vector of each node;

[0010] S5: The spatial embedding vector and the temporal embedding vector are weightedly fused through the cross-spatiotemporal attention mechanism to generate a fused spatiotemporal representation;

[0011] S6: Using the data containing only normal operating samples in the fused spatiotemporal representation, the autoencoder model is trained to learn the low-dimensional reconstruction function of the normal mode;

[0012] S7: Input the spatiotemporal representation collected in real time into the trained autoencoder, calculate the reconstruction error, and map the reconstruction error to a fault risk score based on the threshold determined by the Gaussian mixture model;

[0013] S8: When the risk score exceeds the risk threshold, an early warning message including the fault risk level, suspected fault area and recommended disposal measures is automatically generated and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal.

[0014] In some embodiments, the monitoring data of the distribution network in step S1 includes node voltage, current, active / reactive power, three-phase imbalance and environmental monitoring data; the environmental monitoring data includes temperature and humidity data;

[0015] The data preprocessing steps include: interpolating missing values, filtering outliers and normalizing the acquired data samples.

[0016] In some embodiments, the distribution network topology constructed in step 2 , where V represents the node set in the distribution network topology graph, , is the total number of devices in the constructed distribution network topology; represents the edge set in the distribution network topology graph, , indicating the corresponding node device With node device The physical connection between nodes and feeders in the distribution network topology is established; the mapping relationship between each node and feeder and the actual geographical coordinates is established;

[0017] Among them, each node The corresponding attribute is set to the node The eigenvector of , and each node The eigenvector of Spliced ​​into node feature matrix ;in Represents the dimension of each node's feature vector, that is, the total number of physical quantities used to describe the node state, such as the number of parameters such as voltage, current, active / reactive power, and three-phase imbalance.

[0018] In some embodiments, the multi-head graph attention network in step S3 includes K attention heads, and enhances the spatial feature expression capability by parallel multi-scale neighborhood attention calculation; step S3 specifically includes:

[0019] S3.1: Node feature matrix Each original feature in Apply the same affine transformation in parallel to transform the original features Transform to hidden dimension , the expression is as follows:

[0020]

[0021] in, 、 is the affine transformation parameter; affine transformation can not only reduce the dimension but also provide node representation of the same dimension for subsequent attention operations.

[0022] S3.2: For each edge in the distribution network topology , concatenate the node features after affine transformation and calculate the attention energy , which is expressed as follows:

[0023]

[0024] in, Representation node To its neighbors The energy of attention, represents the learnable weight vector in the attention mechanism, Represents the leakage factor Activation function, used to perform nonlinear mapping of attention energy; Indicated by and The concatenated feature vector;

[0025] S3.3: For the same central node All neighbor scores are normalized into weights according to SoftMax , the sum of the weights of all neighbors is 1; The expression is as follows:

[0026] ,

[0027] in, Representation node The set of neighbor nodes of

[0028] S3.4: Combined with normalized attention weights , weighted and combined node features after affine transformation of neighboring nodes and activation function , the expression is as follows:

[0029]

[0030] Among them, the activation function For ELU, the output is the spatial feature representation under single-head attention;

[0031] S3.5: Each of the K attention heads in the multi-head graph attention network repeats steps S3.1 to S3.4 independently to obtain ;

[0032] For the middle layer, the output of each head is spliced ​​into ; For the final output layer, average the outputs of each head:

[0033]

[0034] by As a node The spatial embedding vector of a node incorporates important features of the node itself and its multi-hop neighbors.

[0035] In some embodiments, the long short-term memory network in step S4 is a bidirectional LSTM; step S4 specifically includes:

[0036] Get Node The pre-processed monitoring features at the current time t and the previous T-1 time constitute the time series matrix : ;

[0037] Every moment The feature is mapped to the latent space to obtain the latent vector at each moment , the expression is as follows:

[0038] , ..., t

[0039] in, 、 is the feature mapping affine transformation parameter, which is used to map the F-dimensional original features to dimensional embedding space; represents a nonlinear activation function;

[0040] will sequence Input bidirectional LSTM to obtain the forward hidden state and backward hidden state respectively, and concatenate them at each moment to obtain the forward and backward hidden state concatenation vector ;

[0041]

[0042] in, represents the forward hidden state, represents the backward hidden state;

[0043] Calculate the attention energy for the hidden state concatenation vector at each moment , and attention energy Perform SoftMax normalization to obtain the moment weight , and then perform weighted aggregation to generate the context vector , The expression is as follows:

[0044]

[0045] Among them, the moment weight The expression is as follows:

[0046]

[0047] Among them, attention energy The expression is as follows:

[0048]

[0049] In the above formula, The linear transformation weight matrix of the attention scoring network is used to transform the bidirectional LSTM at time The spliced ​​hidden state Mapping to the intermediate dimension of attention; is the bias vector of the attention scoring network, and Cooperate to translate and adjust the mapping result; v represents the attention vector, which is used to map the intermediate representation to scalar energy;

[0050] Linearly map the context vector to the same dimension d as the spatial embedding to obtain the temporal embedding vector , The expression is as follows:

[0051]

[0052] in, 、 is a linear mapping parameter used to map the context vector to the d-dimensional space with the same dimension as the embedding vector.

[0053] In some embodiments, step S5 is specifically as follows:

[0054] Embedding space into vectors and time embedding vector Splicing along the feature dimension to form an intermediate representation , then the concatenated vector Apply a two-output unit scoring network to calculate the attention energy of both spatial and temporal paths. The expression is as follows:

[0055]

[0056] in, The linear mapping weight matrix of the cross-temporal and spatial attention scoring network is used to transform the concatenated intermediate representation Mapping to a two-dimensional energy vector; is the bias vector of the cross-temporal attention scoring network, and Cooperate to perform translation correction on the mapping results; 、 are the unnormalized energies of the spatial and temporal components, respectively;

[0057] Energy vector Apply SoftMax operation to obtain the cross-temporal and spatial attention weights, which are expressed as follows:

[0058] ;

[0059] Using the above attention weights, we can sum the spatial and temporal embeddings to get the node The fused spatiotemporal representation of ( ), whose expression is: .

[0060] In some embodiments, in step S6, the autoencoder model is trained using the fused spatiotemporal representation containing only normal operating samples, and the total loss function is minimized by backpropagation through stochastic gradient descent;

[0061] During the training process of the autoencoder model, the decrease of the reconstruction error is monitored on the training set. When the decrease of the error of the validation set is less than the set threshold, the training is stopped and the trained autoencoder is obtained.

[0062] In some embodiments, the Gaussian mixture model (GMM) in step S7 adaptively adjusts the threshold online based on the distribution of historical normal operation reconstruction errors to adapt to seasonal load fluctuations. Step S7 is specifically as follows:

[0063] S7.1: Node The fused spatiotemporal representation at time t Input the trained autoencoder to obtain the reconstruction result ( ), and calculate the node The reconstruction error , which is expressed as follows:

[0064]

[0065] S7.2: For each node Construct and maintain a historical normal error set of length L :

[0066]

[0067] This window only contains the error values ​​during the period when the warning was not triggered, so as to reflect the current season and load conditions;

[0068] S7.3: With the latest As a sample, use the EM algorithm to Fit the Q component to a one-dimensional Gaussian mixture model: , get the node The mixed weight , mean and variance ;

[0069] in, represents the probability density of the Gaussian mixture model; Q is the number of Gaussian components of the Gaussian mixture model, For the Gaussian mixture model q Gaussian components at the nodes ,time t The mixing weights, For the Gaussian mixture model q The mean of the Gaussian components, For the Gaussian mixture model q The standard deviation of the Gaussian components, Indicates the mean , the variance is One-dimensional Gaussian distribution density function;

[0070] S7.4: Based on the preset quantiles , through the cumulative distribution function of the Gaussian mixture model Inverse threshold , which is expressed as follows:

[0071] ;

[0072] in, is the standard normal cumulative distribution function, and the larger the value, the higher the risk;

[0073] S7.5: Map the current reconstruction error to a node Failure risk score :

[0074] , ;

[0075] S7.6: With period Δt as frequency, the latest error Incorporate sliding window , remove the oldest sample, and repeat S7.3 to S7.5 to make the threshold and risk scores Dynamically adjust according to seasonal load fluctuations and network conditions.

[0076] In some embodiments, step S8 specifically includes: constructing a mapping function from risk score to risk level L;

[0077] Establish mapping relationships between various combinations of different risk levels and different area types (such as transformer areas and feeder sections) and a predefined treatment measure library;

[0078] When the risk score exceeds the risk threshold, the following steps are performed:

[0079] A) Based on dynamic threshold Adjust the risk score to generate a second risk score , and determine the second risk score based on the mapping function from risk score to risk level L The corresponding risk level;

[0080] B) All satisfied Node Gathered into a trigger warning node set I, in the distribution network topology diagram In the example, I is the node set, and the corresponding induced subgraph is extracted. ,in, ;

[0081] The induced subgraph is divided into undirected connected components to obtain several node subsets, which are expressed as follows:

[0082]

[0083] Among them, each represents a set of connected nodes, ; represents the set of connected component nodes of mth component; Represents an algorithm for extracting connected components from an undirected graph;

[0084] For each connected component Query the feeder segments corresponding to all its nodes. Based on the mapping relationship between each node, feeder and actual geographic coordinates in the distribution network topology, locate these feeder segments and merge their geographic coordinates to form a closed boundary of the suspected fault area, thereby locating the suspected fault area.

[0085] C) Selecting corresponding treatment measures from the predefined treatment measures library based on the determined risk level and the area type of the suspected fault area, and using them as recommended treatment measures;

[0086] The alarm time, fault risk level, suspected fault area and recommended disposal measures are packaged into a structured early warning message and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal.

[0087] In some embodiments, in step A), based on a dynamic threshold Adjust the risk score to generate a second risk score , specifically including the following steps:

[0088] Calculate the current reconstruction error With dynamic threshold Ratio , and the ratio Perform scale transformation to obtain , the expression is as follows:

[0089]

[0090] in, is the mapping steepness factor, which is used to control the steepness of the mapping curve. , the general value range is [5, 10];

[0091] Next, Input the Sigmoid function to get the final second risk score The second risk score The expression is as follows:

[0092] , ; Used for risk level classification.

[0093] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0094] The technical solution provided by the present invention can realize fault warning of the subway power distribution system, locate the suspected fault area and provide disposal suggestions, thereby ensuring the efficient and stable operation of the subway power distribution system. The spatiotemporal fusion method based on the graph attention network and the autoencoder in the present invention can simultaneously mine the network topology association and the node timing dynamics, significantly improving the recognition accuracy of fault precursors; on the other hand, the autoencoder in the present invention only needs normal operation data to be deployed, does not rely on scarce fault samples, and has the ability to generalize unknown faults; in addition, the Gaussian mixture model can adaptively adjust the threshold according to seasonal load and working condition changes, and then combine the error ratio mapping to ensure the comparability and stability of the risk score in different scenarios; finally, the suspected fault area is accurately located through connected subgraph analysis, and warning information containing the fault risk level, suspected fault area and recommended disposal measures is automatically generated and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal, so that the operation and maintenance personnel can take preventive measures in advance to ensure the efficient and stable operation of the subway power distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0096] Figure 1 This is a schematic diagram of a brief implementation process of the subway power distribution system fault prediction and operation and maintenance method based on big data analysis of the present invention;

[0097] Figure 2 This is an example diagram of the distribution network topology diagram in the solution of the present invention. DETAILED DESCRIPTION

[0098] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0099] See also Figures 1 and 2 This embodiment provides a method for predicting and operating subway power distribution system faults based on big data analysis, including:

[0100] S1: Obtain historical and real-time monitoring data of the distribution network, including node voltage, current, active / reactive power, three-phase imbalance, and environmental monitoring data (including temperature and humidity data), and perform missing value interpolation, outlier filtering, and normalization on the acquired data samples.

[0101] S2: Convert the distribution network topology into a topology graph data model. Map the power equipment (transformers, feeder circuit breakers, etc.) in the distribution network as nodes of the graph according to the physical structure of the distribution network, use feeders as edges of the graph, and assign the pre-processed data to the corresponding node attributes. Specifically, the constructed distribution network topology graph , where V represents the node set in the distribution network topology graph, , is the total number of devices in the constructed distribution network topology; represents the edge set in the distribution network topology graph, , indicating the corresponding node device With node device The physical connection between nodes and feeders in the distribution network topology is established; the mapping relationship between each node and feeder and the actual geographical coordinates is established;

[0102] Among them, each node The corresponding attribute is set to the node The eigenvector of , and each node The eigenvector of Spliced ​​into node feature matrix ; Represents the dimension of each node's feature vector, that is, the total number of physical quantities used to describe the node state, such as the number of parameters such as voltage, current, active / reactive power, and three-phase imbalance.

[0103] S3: Based on the topological graph, a multi-head graph attention network (GAT) is used to aggregate features of the neighborhood of each node to obtain the spatial embedding vector of each node. The GAT includes K attention heads and enhances the spatial feature expression capability through parallel multi-scale neighborhood attention calculation. The specific steps include:

[0104] S3.1: Node feature matrix Each original feature in Apply the same affine transformation in parallel to transform the original features Transform to hidden dimension (such as 64 or 128), the expression is as follows:

[0105]

[0106] in, 、 is the affine transformation parameter; affine transformation can not only reduce the dimension but also provide node representation of the same dimension for subsequent attention operations.

[0107] S3.2: For each edge in the distribution network topology , concatenate the node features after affine transformation and calculate the attention energy , which is expressed as follows:

[0108]

[0109] in, Representation node To its neighbors The energy of attention, represents the learnable weight vector in the attention mechanism, Represents the leakage factor Activation function, used to perform nonlinear mapping of attention energy; Indicated by and The concatenated feature vector;

[0110] S3.3: For the same central node All neighbor scores are normalized into weights according to SoftMax , the sum of the weights of each neighbor is 1, thus reflecting the relative importance distribution; The expression is as follows:

[0111] ,

[0112] in, Representation node The set of neighbor nodes of

[0113] S3.4: Combined with normalized attention weights , weighted and combined node features after affine transformation of neighboring nodes and activation function , the expression is as follows:

[0114]

[0115] Among them, the activation function For ELU, the output is the spatial feature representation under single-head attention;

[0116] S3.5: K attention heads are set in parallel in the multi-head graph attention network. Each attention head repeats steps S3.1 to S3.4 independently to obtain ;

[0117] For the middle layer, the output of each head is spliced ​​into ; For the final output layer, average the outputs of each head:

[0118]

[0119] by As a node The spatial embedding vector of the node is constructed; this vector integrates the important features of the node itself and its multi-hop neighbors, which can reveal the fault propagation situation in a wider range and thus improve the foresight of the early warning.

[0120] S4: Input the monitoring time series data of each node into the long short-term memory network (bidirectional LSTM), extract the time context features, and perform time step aggregation to obtain the time embedding vector of each node. The specific steps include the following:

[0121] Get Node The pre-processed monitoring features at the current time t and the previous T-1 time constitute the time series matrix : ;

[0122] Every moment The feature is mapped to the latent space to obtain the latent vector at each moment , the expression is as follows:

[0123] , ..., t

[0124] in, 、 is the feature mapping affine transformation parameter, which is used to map the F-dimensional original features to dimensional embedding space; Represents a nonlinear activation function, where the nonlinear activation function is ReLU or ELU;

[0125] will sequence Input bidirectional LSTM to obtain the forward hidden state and backward hidden state respectively, and concatenate them at each moment to obtain the forward and backward hidden state concatenation vector ;

[0126]

[0127] in, represents the forward hidden state, Represents the backward hidden state. By simultaneously capturing forward and backward information through a bidirectional LSTM, it can respond to sudden surges in fault precursors while also capturing trend changes over longer periods of time.

[0128] Calculate the attention energy for the hidden state concatenation vector at each moment , and attention energy Perform SoftMax normalization to obtain the moment weight , and then perform weighted aggregation to generate the context vector , The expression is as follows:

[0129]

[0130] Among them, the moment weight The expression is as follows:

[0131]

[0132] Among them, attention energy The expression is as follows:

[0133]

[0134] In the above formula, The linear transformation weight matrix of the attention scoring network is used to transform the bidirectional LSTM at time The spliced ​​hidden state Mapping to the intermediate dimension of attention; is the bias vector of the attention scoring network, and Cooperate to translate and adjust the mapping result; v represents the attention vector, which is used to map the intermediate representation to scalar energy;

[0135] Linearly map the context vector to the same dimension d as the spatial embedding to obtain the temporal embedding vector , The expression is as follows:

[0136]

[0137] in, 、 is a linear mapping parameter used to map the context vector to the same dimension d as the spatial embedding vector. The temporal context vector is unified to the same dimension d as the spatial embedding vector to facilitate subsequent cross-temporal and spatial fusion operations.

[0138] S5: Perform weighted fusion of the spatial embedding vector and the temporal embedding vector through the cross-spatiotemporal attention mechanism to generate a fused spatiotemporal representation. This specifically includes the following steps:

[0139] Embedding space into vectors and time embedding vector Splicing along the feature dimension to form an intermediate representation , then the concatenated vector Apply a two-output unit scoring network to calculate the attention energy of both spatial and temporal paths. The expression is as follows:

[0140]

[0141] in, The linear mapping weight matrix of the cross-temporal and spatial attention scoring network is used to transform the concatenated intermediate representation Mapping to a two-dimensional energy vector; is the bias vector of the cross-temporal attention scoring network, and Cooperate to perform translation correction on the mapping results; 、 are the unnormalized energies of the spatial and temporal components, respectively;

[0142] Energy vector Apply SoftMax operation to obtain the cross-temporal and spatial attention weights, which are expressed as follows:

[0143] ;

[0144] Using the above attention weights, we can sum the spatial and temporal embeddings to get the node The fused spatiotemporal representation of ( ), whose expression is: .

[0145] in, The network topology correlation and temporal dynamics are comprehensively considered, and the contributions of the two are balanced through cross-temporal and spatial attention, providing high-quality spatiotemporal feature representation for subsequent autoencoder training and fault risk scoring, effectively improving the accuracy of distribution network fault warning and real-time response capabilities.

[0146] S6: Using the data containing only normal operating samples in the fused spatiotemporal representation, the autoencoder model is trained to learn the low-dimensional reconstruction function of the normal mode; specifically, the autoencoder model is trained using the fused spatiotemporal representation containing only normal operating samples, and the total loss function is minimized by backpropagation through stochastic gradient descent; during the training of the autoencoder model, the decrease of the reconstruction error is monitored on the training set, and the training is stopped when the decrease of the error on the validation set is less than the set threshold, thereby obtaining a trained autoencoder.

[0147] S7: Input the real-time spatiotemporal representation into the trained autoencoder, calculate the reconstruction error, and map the reconstruction error to a fault risk score based on a threshold determined by a Gaussian mixture model (GMM). The GMM adaptively adjusts the threshold online based on the distribution of reconstruction errors during historical normal operation to accommodate seasonal load fluctuations. This specifically includes the following steps:

[0148] S7.1: Node The fused spatiotemporal representation at time t Input the trained autoencoder to obtain the reconstruction result ( ), and calculate the node The reconstruction error , which is expressed as follows:

[0149]

[0150] S7.2: For each node Construct and maintain a historical normal error set of length L :

[0151]

[0152] This window only contains the error values ​​during the period when the warning was not triggered, so as to reflect the current season and load conditions;

[0153] S7.3: The latest one in memory As a sample, use the EM algorithm to Fit the Q component to a one-dimensional Gaussian mixture model: , get the node The mixed weight , mean and variance ;

[0154] in, represents the probability density of the Gaussian mixture model; Q is the number of Gaussian components of the Gaussian mixture model, For the Gaussian mixture model q Gaussian components at the nodes ,time t The mixing weights, For the Gaussian mixture model q The mean of the Gaussian components, For the Gaussian mixture model q The standard deviation of the Gaussian components, Indicates the mean , the variance is One-dimensional Gaussian distribution density function;

[0155] S7.4: Based on the preset quantiles (e.g. 0.99), through the cumulative distribution function of the Gaussian mixture model Inverse threshold , which is expressed as follows:

[0156] ;

[0157] in, is the standard normal cumulative distribution function (CDF), and a larger value indicates a higher risk;

[0158] S7.5: Map the current reconstruction error to a node Failure risk score :

[0159] , ;

[0160] S7.6: With a period Δt (e.g., 1 h), update the latest error Incorporate sliding window , remove the oldest sample, and repeat S7.3 to S7.5 to make the threshold and risk scores Dynamically adjust to seasonal load fluctuations and network conditions. This solution directly utilizes historical normal reconstruction error data, rather than relying on manually set fixed thresholds. It automatically calibrates thresholds based on actual operating conditions (such as seasonal load fluctuations and network topology adjustments). Compared to static thresholds, this significantly reduces missed and false alarm rates, eliminating the need for frequent parameter readjustment by operations and maintenance personnel.

[0161] S8: When the risk score exceeds the risk threshold, an early warning message including the fault risk level, suspected fault area, and recommended treatment measures is automatically generated and sent to at least one power maintenance personnel terminal or grid operation monitoring terminal; specifically, the following steps are included:

[0162] Construct a mapping function from risk score to risk level L. In this embodiment, the risk level is determined by setting a multi-stage risk threshold. For example, setting a three-stage risk threshold ,in and When the risk score ≤ When , it is judged as low risk; when ≤ When ≤ When the risk score is > It is considered an emergency state.

[0163] A mapping relationship is established between various combinations of different risk levels and different area types (such as transformer substations, feeder sections, etc.) (for example, low risk + transformer substations, medium risk + transformer substations, etc.) and a predefined disposal measure library; the predefined disposal measure library contains disposal measures for different combinations of risk levels and different area types, for example: the disposal measure corresponding to "low risk + transformer substation" is "check alarm records", and the disposal measure corresponding to "medium risk + transformer substation" is "on-site inspection", etc.

[0164] When the risk score exceeds the risk threshold ( ), then execute the following steps:

[0165] A) Based on dynamic threshold Adjust the risk score to generate a second risk score In this embodiment, the second risk score The generation method is as follows:

[0166] Calculate the current reconstruction error With dynamic threshold Ratio , and the ratio Perform scale transformation to obtain , the expression is as follows:

[0167]

[0168] in, is the mapping steepness factor, which is used to control the steepness of the mapping curve. ( The general value range is [5, 10]);

[0169] Next, Input the Sigmoid function to get the final second risk score ; Second risk score The expression is: , ; Can be directly used for subsequent risk level classification. Based on dynamic threshold Adjusting the risk scores helps maintain consistency in risk classification under different working conditions.

[0170] Then, according to the mapping function from risk score to risk level L, the second risk score is determined The corresponding risk level.

[0171] B) All satisfied Node Gathered into a trigger warning node set I, in the distribution network topology diagram In the example, I is the node set, and the corresponding induced subgraph is extracted. ,in, ;

[0172] The induced subgraph is divided into undirected connected components to obtain several node subsets, which are expressed as follows:

[0173]

[0174] Among them, each represents a set of connected nodes, ; represents the set of connected component nodes of mth component; Represents an algorithm for extracting connected components from an undirected graph;

[0175] For each connected component The system queries the feeder segments corresponding to all nodes. Based on the mapping relationship between each node and feeder in the distribution network topology and the actual geographic coordinates, these feeder segments are located and their geographic coordinates are merged to form a closed boundary of the suspected fault area, thereby locating the suspected fault area. This method can automatically aggregate several triggering nodes into one or more connected areas based on the set of nodes that triggered the warning and the distribution network topology. This avoids repeated and discrete alarms for individual nodes, helps operation and maintenance personnel identify key areas of concern, and improves the efficiency of on-site surveys and emergency repairs.

[0176] C) Based on the determined risk level and the area type of the suspected fault area (such as transformer substation, feeder section, etc.), the corresponding treatment measures are selected from the predefined treatment measures library and used as recommended treatment measures; the alarm time, fault risk level, suspected fault area and recommended treatment measures are packaged into a structured early warning message and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal.

[0177] Based on the above, the solution of the present invention significantly improves the accuracy and sensitivity of distribution network fault warnings through a full-link design from network topology, timing dynamics to adaptive fusion, dynamic threshold adjustment, and regionalized early warning, reduces operation and maintenance costs, and ensures the efficient and stable operation of the subway distribution system.

[0178] In addition, the functional units in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0179] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0180] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for fault prediction and operation and maintenance of subway power distribution system based on big data analysis, characterized in that: include: S1: Obtain historical and real-time monitoring data of the distribution network and preprocess the acquired data samples; S2: Convert the distribution network topology into a topology graph data model. Map each power device in the distribution network as a node of the graph based on the physical structure of the distribution network, with feeders as edges of the graph, and assign pre-processed data to corresponding node attributes. S3: Based on the topological graph, a multi-head graph attention network is used to aggregate features of the neighborhood of each node to obtain the spatial embedding vector of each node; S4: Input the monitoring time series data of each node into the long short-term memory network, extract the time context features, and perform time step aggregation to obtain the time embedding vector of each node; S5: The spatial embedding vector and the temporal embedding vector are weightedly fused through the cross-spatiotemporal attention mechanism to generate a fused spatiotemporal representation; specifically: Embedding space into vectors and time embedding vector Splicing along the feature dimension to form an intermediate representation , then the concatenated vector Apply a two-output unit scoring network to calculate the attention energy of both spatial and temporal paths. The expression is as follows: in, The linear mapping weight matrix of the cross-temporal and spatial attention scoring network is used to transform the concatenated intermediate representation Mapping to a two-dimensional energy vector; is the bias vector of the cross-temporal attention scoring network, and Cooperate to perform translation correction on the mapping results; 、 are the unnormalized energies of the spatial and temporal components, respectively; Energy vector Apply SoftMax operation to obtain the cross-temporal and spatial attention weights, which are expressed as follows: ； Using the above attention weights, we can sum the spatial and temporal embeddings to get the node The fused spatiotemporal representation of , whose expression is: ; S6: Using the data containing only normal operating samples in the fused spatiotemporal representation, the autoencoder model is trained to learn the low-dimensional reconstruction function of the normal mode; S7: Input the spatiotemporal representation collected in real time into the trained autoencoder, calculate the reconstruction error, and map the reconstruction error to a fault risk score based on the threshold determined by the Gaussian mixture model; S8: When the risk score exceeds the risk threshold, an early warning message including the fault risk level, suspected fault area and recommended disposal measures is automatically generated and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal.

2. The method for predicting and operating subway power distribution system faults based on big data analysis according to claim 1 is characterized in that: The monitoring data of the distribution network in step S1 includes node voltage, current, active / reactive power, three-phase imbalance and environmental monitoring data; the environmental monitoring data includes temperature and humidity data; The data preprocessing steps include: interpolating missing values, filtering outliers and normalizing the acquired data samples.

3. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 1 is characterized in that: The distribution network topology constructed in step 2 , where V represents the node set in the distribution network topology graph, , is the total number of devices in the constructed distribution network topology; represents the edge set in the distribution network topology graph, , indicating the corresponding node device With node device The physical connection between nodes and feeders in the distribution network topology is established; the mapping relationship between each node and feeder and the actual geographical coordinates is established; Among them, each node The corresponding attribute is set to the node The eigenvector of , and each node The eigenvector of Spliced ​​into node feature matrix ;in Represents the dimension of each node feature vector.

4. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 3 is characterized in that: The multi-head graph attention network in step S3 includes K attention heads, which enhances the spatial feature expression capability by parallel multi-scale neighborhood attention calculation; step S3 specifically includes: S3.1: Node feature matrix Each original feature in Apply the same affine transformation in parallel to transform the original features Transform to hidden dimension , the expression is as follows: in, 、 is the affine transformation parameter; S3.2: For each edge in the distribution network topology , concatenate the node features after affine transformation and calculate the attention energy , which is expressed as follows: in, Representation node To its neighbors The energy of attention, represents the learnable weight vector in the attention mechanism, Represents the leakage factor Activation function, used to perform nonlinear mapping of attention energy; Indicated by and The concatenated feature vector; S3.3: For the same central node All neighbor scores are normalized into weights according to SoftMax , the sum of the weights of all neighbors is 1; The expression is as follows: ， in, Representation node The set of neighbor nodes of S3.4: Combined with normalized attention weights , weighted and combined node features after affine transformation of neighboring nodes and activation function , the expression is as follows: Among them, the activation function For ELU, the output is the spatial feature representation under single-head attention; S3.5: Each of the K attention heads in the multi-head graph attention network repeats steps S3.1 to S3.4 independently to obtain ; For the middle layer, the output of each head is spliced ​​into ; For the final output layer, average the outputs of each head: by As a node The spatial embedding vector of .

5. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 3 is characterized in that: The long short-term memory network in step S4 is a bidirectional LSTM; step S4 specifically includes: Get Node The pre-processed monitoring features at the current time t and the previous T-1 time constitute the time series matrix : ; Every moment The feature is mapped to the latent space to obtain the latent vector at each moment , the expression is as follows: ， ...， t in, 、 is the feature mapping affine transformation parameter, which is used to map the F-dimensional original features to dimensional embedding space; represents a nonlinear activation function; will sequence Input bidirectional LSTM to obtain the forward hidden state and backward hidden state respectively, and concatenate them at each moment to obtain the forward and backward hidden state concatenation vector ; in, represents the forward hidden state, represents the backward hidden state, Represents vector concatenation operation; Calculate the attention energy for the hidden state concatenation vector at each moment , and attention energy Perform SoftMax normalization to obtain the moment weight , and then perform weighted aggregation to generate the context vector , The expression is as follows: Among them, the moment weight The expression is as follows: Among them, attention energy The expression is as follows: In the above formula, The linear transformation weight matrix of the attention scoring network is used to transform the bidirectional LSTM at time The spliced ​​hidden state Mapping to the intermediate dimension of attention; is the bias vector of the attention scoring network, and Cooperate to translate and adjust the mapping result; v represents the attention vector, which is used to map the intermediate representation to scalar energy; Linearly map the context vector to the same dimension d as the spatial embedding to obtain the temporal embedding vector , The expression is as follows: in, 、 is a linear mapping parameter used to map the context vector to the d-dimensional space with the same dimension as the embedding vector.

6. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 3 is characterized in that: In step S6, the autoencoder model is trained using the fused spatiotemporal representation containing only normal running samples, and the total loss function is minimized by backpropagation through stochastic gradient descent; During the training process of the autoencoder model, the decrease of the reconstruction error is monitored on the training set. When the decrease of the error of the validation set is less than the set threshold, the training is stopped and the trained autoencoder is obtained.

7. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 3 is characterized in that: The Gaussian mixture model in step S7 adaptively adjusts the threshold online based on the distribution of historical normal operation reconstruction errors to adapt to seasonal load fluctuations. Step S7 is specifically as follows: S7.1: Node The fused spatiotemporal representation at time t Input the trained autoencoder to obtain the reconstruction result , and calculate the node The reconstruction error , which is expressed as follows: S7.2: For each node Construct and maintain a historical normal error set of length L : Should The window only includes the error values ​​during the period when the warning was not triggered, so as to reflect the current season and load conditions; S7.3: With the latest As a sample, use the EM algorithm to Fit the Q component to a one-dimensional Gaussian mixture model: , get the node The mixed weight , mean and variance ; in, represents the probability density of the Gaussian mixture model; Q is the number of Gaussian components of the Gaussian mixture model, For the Gaussian mixture model q Gaussian components at the nodes ,time t The mixing weights, For the Gaussian mixture model q The mean of the Gaussian components, For the Gaussian mixture model q The standard deviation of the Gaussian components, Indicates the mean , the variance is One-dimensional Gaussian distribution density function; S7.4: Based on the preset quantiles , through the cumulative distribution function of the Gaussian mixture model Inverse threshold , which is expressed as follows: ； in, is the standard normal cumulative distribution function, and the larger the value, the higher the risk; S7.5: Map the current reconstruction error to a node Failure risk score : ， ； S7.6: With period Δt as frequency, the latest error Incorporate sliding window , remove the oldest sample, and repeat S7.3 to S7.5 to make the threshold and risk scores Dynamically adjust according to seasonal load fluctuations and network conditions.

8. The method for subway power distribution system fault prediction and operation and maintenance based on big data analysis according to claim 7 is characterized in that: Step S8 specifically includes: constructing a mapping function from risk score to risk level L; Establish mapping relationships between various combinations of different risk levels and different area types and a predefined disposal measures library; When the risk score exceeds the risk threshold, the following steps are performed: A) Based on dynamic threshold Adjust the risk score to generate a second risk score , and determine the second risk score based on the mapping function from risk score to risk level L The corresponding risk level; B) All satisfied Node Gathered into a trigger warning node set I, in the distribution network topology diagram In the example, I is the node set, and the corresponding induced subgraph is extracted. ,in, ; The induced subgraph is divided into undirected connected components to obtain several node subsets, which are expressed as follows: Among them, each represents a set of connected nodes, ; represents the set of connected component nodes of mth component; Represents an algorithm for extracting connected components from an undirected graph; For each connected component Query the feeder segments corresponding to all its nodes. Based on the mapping relationship between each node, feeder and actual geographic coordinates in the distribution network topology, locate these feeder segments and merge their geographic coordinates to form a closed boundary of the suspected fault area, thereby locating the suspected fault area. C) Selecting corresponding treatment measures from the predefined treatment measures library based on the determined risk level and the area type of the suspected fault area, and using them as recommended treatment measures; The alarm time, fault risk level, suspected fault area and recommended disposal measures are packaged into a structured early warning message and sent to at least one power maintenance personnel terminal or power grid operation monitoring terminal.

9. The subway power distribution system fault prediction and operation and maintenance method based on big data analysis according to claim 8 is characterized in that: In step A), the dynamic threshold Adjust the risk score to generate a second risk score , specifically including the following steps: Calculate the current reconstruction error With dynamic threshold Ratio , and the ratio Perform scale transformation to obtain , the expression is as follows: in, is the mapping steepness factor, which is used to control the steepness of the mapping curve. ; Next, Input the Sigmoid function to get the final second risk score The second risk score The expression is as follows: ， 。

Images