Method for realizing fault prediction based on power distribution network space diagram

Abstract

The invention discloses a method for realizing fault prediction based on a power distribution network space diagram. The method comprises the following steps: acquiring original multi-source data to generate multi-source data; performing data format conversion on the multi-source data to generate multi-dimensional time series data, wherein the data format of the multi-dimensional time series data is graph structure data + time series data; constructing a power distribution network space diagram model to obtain a space feature vector; the power distribution network space graph model is represented by an adjacency matrix A and a node feature matrix X, the adjacency matrix A reflects spatial features between nodes, the node feature matrix X reflects node features, and a spatial feature vector is used for judging a fault risk value of each node; generating a power distribution network fault space distribution diagram by combining the power distribution network space diagram; and fusing the spatial features and the time features, and realizing fault prediction based on the power distribution network spatial diagram. According to the technical scheme, accurate positioning of the fault area can be realized, positioning errors are reduced, high-risk area identification accuracy is improved, fault prediction accuracy is improved, and dynamic adjustment of regional geographical environment and meteorological characteristics is adapted.

Description

Technical Field

[0001] This invention relates to the intersection of power system operation and maintenance and artificial intelligence, and more specifically, to a method for fault prediction based on a distribution network spatial map. Background Technology

[0002] The power distribution network is the final link in the power system, directly supplying power to users, and its operation is closely related to social production and daily life. The distribution network is characterized by its wide distribution of lines, complex geographical environment, and diverse equipment types. It is susceptible to multiple factors such as extreme weather (heavy rain, lightning, icing), terrain conditions (mountains, vegetation cover), and equipment aging, resulting in frequent faults that exhibit significant spatial heterogeneity and temporal randomness. Once a fault occurs, if the fault area cannot be quickly located and early warnings issued, the power outage area will expand, repair efficiency will be low, and significant economic losses and social impacts will result.

[0003] Currently, fault mapping in distribution network fault management relies on static GIS maps to mark historical fault points. This method lacks dynamic integration of real-time equipment status and environmental factors, failing to intuitively present the spatial distribution patterns of faults and potential risk areas. The mapping results are lagging and have limited practicality. On the other hand, fault prediction methods often focus on time-series analysis in a single time dimension (such as LSTM capturing only time dependencies), ignoring the spatial correlation characteristics of the distribution network topology. This makes it difficult to effectively uncover the fault evolution patterns under the multi-dimensional coupling of "space-time-environment". Furthermore, multi-source data (power grid topology, sensor data, meteorological and geographical data) suffers from heterogeneous formats and complex correlations. Existing technologies lack efficient fusion mechanisms, resulting in insufficient data value extraction and prediction accuracy and mapping accuracy failing to meet actual operation and maintenance needs.

[0004] Preliminary attempts to combine GIS and machine learning have emerged in existing technologies, but there are still obvious shortcomings: First, spatial feature extraction is insufficient, and no special model has been designed for the graph structure characteristics of the distribution network topology, making it impossible to effectively capture the spatial dependencies between devices; second, mapping and prediction are disconnected, and spatial distribution visualization is not dynamically linked with time-series prediction results, making it difficult to support the entire process of "early warning-location-disposal" operation and maintenance; third, dynamic adaptability is poor, and it is impossible to update mapping results and prediction models according to real-time data, making it difficult to cope with rapidly changing operating environments.

[0005] Therefore, the industry urgently needs an intelligent technology that can integrate multi-source heterogeneous data, deeply mine spatiotemporal coupling characteristics, and achieve the integration of "spatial visualization and time series prediction". Summary of the Invention

[0006] To achieve the above objectives, this application provides a method for fault prediction based on a distribution network spatial map, comprising the following steps: Collect raw multi-source data; preprocess the raw multi-source data to generate multi-source data; the multi-source data includes power grid topology data and equipment location data, real-time sensor monitoring data, meteorological and geographical data, and historical fault data. Multi-source data is converted into multi-dimensional time-series data. The data format of the multi-dimensional time-series data is: graph structure data + time-series sequence data. The graph structure data consists of nodes and edges with attributes, where nodes are devices / buses and edges are line connections. The attributes of the edges include line parameters and environmental characteristics. The time-series sequence data is a multi-dimensional time-series matrix aligned to timestamps. A spatial graph model of the power distribution network is constructed to obtain spatial feature vectors. The spatial graph model of the power distribution network is represented by an adjacency matrix A and a node feature matrix X. The adjacency matrix A reflects the spatial characteristics that express the relationship between each node, and the node feature matrix X reflects the node characteristics. The spatial feature vectors are used to determine the fault risk value of each node. Generate a spatial distribution map of distribution network faults by combining the distribution network spatial map; By integrating spatial and temporal characteristics, fault prediction of nodes can be achieved based on the spatial map of the distribution network.

[0007] The process of constructing a spatial graph model of a power distribution network to obtain spatial feature vectors includes the following steps: Construct a two-dimensional spatial model of the power distribution network and overlay partial content from multiple sources of data to generate a basic spatial map of the power distribution network. Based on the two-dimensional spatial model of the power distribution network, a spatial graph model of the power distribution network is constructed according to graph structure data; the graph structure data refers to the data associated with the power grid lines. GCN is used to learn the spatial dependencies of nodes in the spatial graph model of the distribution network, and the local and global spatial feature vectors of the nodes are extracted as spatial dependencies and spatial feature vectors.

[0008] Among them, node characteristics include the device's own attributes, the environmental parameters of the area where the node is located, and the historical fault attributes corresponding to the node; Spatial characteristics include topological association characteristics, spatial proximity influence characteristics, and regional common characteristics. Among them, topological association characteristics include the connection strength between nodes and adjacent devices and the fault risk transmission coefficient of adjacent nodes; spatial proximity influence characteristics include the equipment status of surrounding nodes and the spatial consistency of the regional environment; regional common characteristics include the average equipment aging degree within the same topological partition, the spatial coverage of regional meteorological disasters, and fault susceptibility clustering caused by topography.

[0009] Furthermore, generating a spatial distribution map of distribution network faults refers to: integrating spatial risk levels with the spatial map of the distribution network, using different colors to mark regional risks, and generating a spatial distribution map of distribution network faults. It receives real-time sensor monitoring data, meteorological and geographical data, and updates the spatial risk level and distribution network fault spatial distribution map according to a certain period.

[0010] Furthermore, node failure prediction refers to outputting the probability of failure for each node / line within the next 1 to 24 hours using the GCN-BiLSTM ensemble model.

[0011] The construction of the GCN-BiLSTM ensemble model includes the following steps: Obtain spatial feature vectors and multidimensional time series data of all nodes; extract timestamps and corresponding meteorological data from the multidimensional time series data as feature sets, extract timestamps and corresponding historical fault attributes as label sets, and the feature sets and label sets constitute the fault prediction dataset. Define the GCN-BiLSTM ensemble model, which includes an input layer, a GCN feature enhancement layer, a BiLSTM temporal modeling layer, and an output layer. The fault prediction dataset is divided into a training set, a validation set, and a test set. The training set is used for parameter learning of the GCN-BiLSTM ensemble model, the validation set is used for hyperparameter tuning of the GCN-BiLSTM ensemble model, and the test set is used for performance evaluation of the GCN-BiLSTM ensemble model. Model evaluation and optimization were performed to complete the construction of the GCN-BiLSTM ensemble model.

[0012] The input layer concatenates spatial feature vectors and multidimensional temporal data to form a "space-time" joint input matrix. The dimensions of the "space-time" joint input matrix include: time step T, number of nodes N, and number of features F, where T=24 and F≥12; the features corresponding to the number of features F cover node features and spatial features.

[0013] Further preprocessing includes outlier detection and removal, missing value imputation, and data normalization; Among them, outlier detection and removal refers to setting an outlier threshold based on the characteristics of distribution network data, and identifying and removing sensor data and logically contradictory topology data that exceed the outlier threshold; missing value filling refers to using graph attention-based interpolation methods to fill missing spatiotemporal data by utilizing the spatial correlation of adjacent nodes and historical time series data; data normalization refers to using Z-Score normalization to normalize numerical data and map the data to the [-1,1] interval.

[0014] When performing data format conversion, multi-source data are classified: data associated with power grid lines is treated as graph structure data; data associated with timestamps is treated as time series data.

[0015] This invention addresses the problem that existing static mapping methods cannot integrate equipment status and environmental data in real time, making it difficult to reflect the dynamic changes in fault spatial risks. It achieves real-time updates of fault spatial distribution maps through the collaboration of GIS and edge computing, intuitively presenting the distribution of risk areas. It mines spatial features using graph neural networks and captures temporal evolution patterns using time-series models, achieving accurate predictions under spatiotemporal coupling. A dedicated data fusion mechanism is employed to achieve unified representation and correlation mining of heterogeneous data, providing high-quality data support for mapping and prediction. Finally, a distribution network spatial map model is constructed to generate a distribution network spatial map, providing spatial features for the prediction model. The prediction results feed back into a linkage mechanism that dynamically updates the distribution network spatial map model, achieving integrated intelligent decision-making. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a method for fault prediction based on a distribution network spatial map, according to an embodiment of the present invention. Detailed Implementation

[0017] This invention uses the GCN-BiLSTM spatiotemporal joint model to adapt to the topological characteristics of the distribution network diagram, while capturing the spatial dependence and temporal evolution laws to achieve visualized fault spatial distribution and fault prediction.

[0018] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] The method for fault prediction based on distribution network spatial diagrams provided by this invention is as follows: Figure 1 As shown, it includes the following steps: Step S100: Collect raw multi-source data; preprocess the raw multi-source data to generate multi-source data; First, the following four types of core data are collected as raw multi-source data through a distributed data acquisition system to provide a foundation for spatial mapping and fault prediction: 1) Power grid topology data and equipment location data, including the location coordinates of distribution network nodes (equipment, busbars), line connection relationships, line parameters (impedance, length), equipment models and years of operation. This type of data is obtained from the power grid GIS system and equipment management platform at a certain frequency (e.g., once a day).

[0020] 2) Real-time sensing and monitoring data, including data such as voltage, current, temperature, and partial discharge signals collected by sensors deployed on lines and equipment, and data sampling frequency set (e.g., 1 minute / time); at the same time, equipment status signals (normal / abnormal) are collected and uploaded to edge devices in real time.

[0021] 3) Meteorological and geographic data, including regional average temperature (-40℃~60℃), relative humidity (0%~100%), wind speed (0~60m / s), rainfall intensity (0~200mm / h), and lightning activity frequency, which are obtained in real time through meteorological department APIs, with an update frequency of 15 minutes / time; geographic data includes terrain type (mountain / plain / hill), vegetation coverage (0%~100%), and altitude (-50m~5000m) obtained through GIS geographic databases.

[0022] 4) Historical fault data, including the time of occurrence, geographical location, fault type (short circuit / grounding / equipment damage), fault cause, and handling result of historical faults obtained from the equipment management platform, with a data time span of no less than 5 years.

[0023] The raw multi-source data collected above is heterogeneous data. To ensure data quality, the raw multi-source data is preprocessed, including outlier detection and removal, missing value imputation, and data normalization. 1) Outlier detection and removal: An improved isolated forest algorithm is adopted, and an anomaly threshold is set in combination with the characteristics of distribution network data. Sensor data exceeding the anomaly threshold (such as 3 times the standard deviation) and logically contradictory topology data are identified and removed. After outlier detection and removal, the accuracy of the original multi-source data can be ≥95%.

[0024] 2) Missing value imputation: For spatiotemporally missing data, a graph attention-based interpolation method is used to imput the missing values ​​by utilizing the spatial correlation between adjacent nodes and historical time series data.

[0025] 3) Data Normalization: For numerical data (voltage, current, temperature, humidity, etc.), Z-Score normalization is applied to map the data to the [-1,1] interval, eliminating dimensional differences; the normalization method is as follows: , where μ is the mean and σ is the standard deviation.

[0026] After preprocessing, the raw multi-source data becomes accurate and complete multi-source data.

[0027] Step S110: Convert the data format of the multi-source data to generate multi-dimensional time series data; The multi-source data is classified as follows: data associated with power grid lines (such as equipment / buses, lines) is treated as graph structure data; data associated with timestamps is treated as time series data; the multi-source data is further converted into a format to generate multi-dimensional time series data; the data format of multi-dimensional time series data is: graph structure data + time series sequence data; Graph structure data consists of nodes and edges with attributes, where nodes are devices / buses and edges are line connections. The attributes of the edges include line parameters and environmental characteristics. Time series data are multidimensional time series matrices aligned to timestamps.

[0028] Step S120: Construct a spatial diagram model of the distribution network and obtain spatial feature vectors; including the following steps: 1) Construct a two-dimensional spatial model of the power distribution network based on a geographic information system (GIS). Overlay the power grid topology data, equipment location data, and geographic data (two-dimensional geographic information such as terrain and vegetation distribution) from multiple sources onto the two-dimensional spatial model of the power distribution network to generate a basic spatial map of the power distribution network. The basic spatial map of the power distribution network can be zoomed, panned, and displayed in layers, including the line layer, equipment layer, and geographic background layer, which can intuitively present the planar relationship between the line direction, equipment distribution, and geographic environment.

[0029] 2) Based on the two-dimensional spatial model of the distribution network, a spatial graph model of the distribution network is constructed according to the graph structure data, including: constructing a spatial graph model of the distribution network with equipment / bus as nodes and lines as edges; the spatial graph model of the distribution network is represented by the adjacency matrix A and the node feature matrix X. The adjacency matrix A reflects the spatial characteristics of the relationship between each node, and the node feature matrix X reflects the node characteristics.

[0030] The node characteristics include the device's own attributes, the environmental parameters of the area where the node is located, and the historical fault attributes corresponding to the node. The device's own attributes include: device model, years of operation, line parameters (impedance, length), real-time status signals (normal / abnormal), and sensor monitoring data (voltage, current, temperature, partial discharge). The regional environmental parameters include meteorological data (temperature, humidity, wind speed, rainfall intensity, number of lightning strikes) and geographical data (terrain type, vegetation coverage, altitude) of the node's location. The historical fault attributes include the historical fault frequency, fault type, and fault cause corresponding to the node.

[0031] Spatial characteristics include topological association characteristics, spatial proximity influence characteristics, and regional common characteristics. Among them, topological association characteristics include the connection strength between nodes and adjacent equipment (such as line impedance and length) and the fault risk transmission coefficient of adjacent nodes (such as the association weight between critical busbars and branch equipment). Spatial proximity influence characteristics include the equipment status of surrounding nodes (such as the spread of abnormal signals from adjacent equipment) and the spatial consistency of the regional environment (such as the environmental synergistic influence of multiple nodes within the same rainfall zone). Regional common characteristics include the average equipment aging degree within the same topological partition, the spatial coverage of regional meteorological disasters, and the clustering of fault susceptibility caused by topography (such as the common risk characteristics of multiple nodes in mountainous areas).

[0032] In general, node characteristics are the "independent attributes" of a single device / region, which do not involve the relationship with other devices and are the basis for constructing spatial characteristics; spatial characteristics, on the other hand, take node characteristics as the core and superimpose the "relationships" of the distribution network topology, which are the integrated characteristics of "individual attributes + group relationships".

[0033] An example in the distribution network spatial diagram model, namely the distribution network spatial diagram.

[0034] 3) Use GCN (Graph Convolutional Network) to learn the spatial dependencies of nodes in the spatial graph model of the distribution network, and extract the local and global spatial feature vectors of the nodes as spatial dependencies, represented as: Where à is the normalized adjacency matrix, H (l) W represents the hidden features (i.e., spatial feature vectors) of the l-th layer. (l) This is the weight matrix. σ This is the Sigmoid activation function.

[0035] 4) Determine the fault risk value of each node based on the spatial feature vector.

[0036] Spatial feature vectors are the core basis for judging the fault risk value of each node. By encoding "equipment's own state + surrounding environment + topology dependence", they directly output feature representations that can be used for risk classification. The dimension and strength of the spatial feature vector correspond to the level of fault risk. Spatial feature vectors dynamically capture the coupling effect of "space-equipment-environment", upgrading fault risk assessment from single-equipment judgment to regional linkage judgment. The stronger the signal of "abnormal equipment association, adverse environment superposition, and topological key node" reflected in the spatial feature vector, the higher the fault risk value, and the more closely it fits the spatial correlation law of distribution network faults.

[0037] Based on the node characteristics and spatial characteristics of the learning distribution network spatial graph model, the fault risk value corresponding to the node characteristics reflects whether the node has potential faults, while the fault risk value corresponding to the spatial characteristics reflects the overall fault risk in the power grid topology after being affected by surrounding equipment / environment. In this step, spatial risk levels are divided according to the risk value as follows: low risk (0~0.3), medium risk (0.3~0.7), and high risk (0.7~1.0).

[0038] Step S130: Generate a spatial distribution map of distribution network faults by combining the distribution network spatial map; The spatial risk level generated in step S120 is integrated with the distribution network spatial map, and different colors (green - low risk, yellow - medium risk, red - high risk) are used to mark the regional risks to generate a distribution network fault spatial distribution map.

[0039] Based on the data acquisition system, real-time sensor monitoring data, meteorological and geographical data are received and updated at certain intervals (such as every 30 minutes) to update the spatial risk level and distribution network fault spatial distribution map, ensuring that the distribution network fault spatial distribution map can reflect the current real-time operating status of the power grid.

[0040] Step S140: Integrate spatial and temporal features to predict node faults based on the distribution network spatial map.

[0041] In this step, the failure probability of each node / line in the next 1 to 24 hours is output through the GCN-BiLSTM ensemble model.

[0042] The steps to construct the GCN-BiLSTM ensemble model are as follows: 1) Obtain the spatial feature vectors and multidimensional time series data of all nodes; extract the timestamps and corresponding meteorological data from the multidimensional time series data as feature sets, and extract the timestamps and corresponding historical fault attributes as label sets. The feature sets and label sets constitute the fault prediction dataset. 2) Define the GCN-BiLSTM ensemble model, which includes an input layer, a GCN feature enhancement layer, a BiLSTM temporal modeling layer, and an output layer; The input layer concatenates spatial feature vectors and multidimensional time series data to form a "space-time" joint input matrix. The dimensions of the "space-time" joint input matrix are (time step T, number of nodes N, number of features F), where T = 24 (hours) and F ≥ 12. The features corresponding to the number of features F cover both node features and spatial features.

[0043] The GCN feature enhancement layer is used to extract features from the spatial dimension of the "space-time" joint input matrix again, thereby strengthening the spatial correlation representation between devices. The BiLSTM time-series modeling layer processes time-series data from both forward and backward directions using a bidirectional long short-term memory network, capturing the time dependencies of fault evolution, represented as: Forward LSTM: , Inverse LSTM: , Fusion Output: .

[0044] The output layer outputs the probability of each node / line failing within the next 124 hours (0: no failure, 1: failure) through a fully connected layer and a sigmoid activation function.

[0045] 3) Train and generate the GCN-BiLSTM ensemble model: The fault prediction dataset is divided into a training set, a validation set, and a test set. The training set is used for parameter learning of the GCN-BiLSTM ensemble model, the validation set is used for hyperparameter tuning of the GCN-BiLSTM ensemble model, and the test set is used for performance evaluation of the GCN-BiLSTM ensemble model.

[0046] Specifically, the training parameters are set as follows: the Adam optimizer is used, the initial learning rate is 0.001, and the learning rate is dynamically adjusted using a cosine annealing strategy; the loss function is the cross-entropy loss function, the number of training iterations is 100 rounds, and the batch size is 32; a Dropout layer (dropout rate=0.3) is added between the GCN layer and the BiLSTM layer, and L2 regularization (λ=0.001) is used to suppress overfitting and improve the model's generalization ability.

[0047] 4) Conduct model evaluation and optimization: Furthermore, accuracy (Acc), recall (Rec), F1 score, and AUC score were used as evaluation metrics to evaluate the GCN-BiLSTM ensemble model. The final evaluation results on the test set were: Acc ≥ 90%, Rec ≥ 85%, F1 ≥ 85%, and AUC ≥ 0.90.

[0048] At this point, the GCN-BiLSTM ensemble model is complete. By acquiring meteorological data for each region over the next 1 to 24 hours from the meteorological system, it is possible to predict whether a fault will occur in each region within the next 1 to 24 hours. Furthermore, by combining the spatial distribution map of the power distribution network faults, high-probability fault areas (probability ≥ 0.6) can be marked on the GIS map to generate an integrated prediction report of "spatial distribution-temporal early warning".

[0049] Meanwhile, based on the federated learning architecture, when information such as power grid topology data, equipment location data, and geographical data changes, new data is collected in a timely manner and the model is trained, thereby improving the model's adaptability to different regions and environments while protecting data privacy.

[0050] This invention extracts spatial correlation features of the power distribution network through GCN and combines them with dynamic visualization via GIS to achieve accurate location of fault areas, greatly reducing location errors and improving the accuracy of high-risk area identification, providing precise spatial guidance for emergency repairs. Simultaneously, it employs a GCN-BiLSTM model to fuse spatiotemporal dual-dimensional features, improving fault prediction accuracy compared to traditional single-time-series models (such as LSTM). This invention solves the problem of heterogeneous data integration through a multi-source data fusion mechanism, fully exploring fault patterns under the coupling of "power grid-environment-spatiotemporal." Furthermore, it supports edge computing and federated learning, dynamically adjusting the model according to different regional geographical environments and meteorological characteristics, maintaining high reliability in various scenarios such as mountains, plains, and coastal areas, and has a wide range of applications.

[0051] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for fault prediction based on a distribution network spatial map, characterized in that, Includes the following steps: Collect raw multi-source data; preprocess the raw multi-source data to generate multi-source data; the multi-source data includes power grid topology data and equipment location data, real-time sensor monitoring data, meteorological and geographical data, and historical fault data. The multi-source data is converted into multi-dimensional time-series data through data format conversion. The data format of the multidimensional time series data is: graph structure data + time series data; wherein, the graph structure data consists of nodes and edges with attributes, where nodes are devices / buses and edges are line connections, and the attributes of the edges include line parameters and environmental characteristics; the time series data is a multidimensional time series matrix aligned by timestamps; A spatial graph model of the power distribution network is constructed to obtain spatial feature vectors. The spatial graph model of the power distribution network is represented by an adjacency matrix A and a node feature matrix X. The adjacency matrix A reflects the spatial characteristics that express the relationship between each node, and the node feature matrix X reflects the node characteristics. The spatial feature vectors are used to determine the fault risk value of each node. Generate a spatial distribution map of distribution network faults by combining the distribution network spatial map; By integrating spatial and temporal characteristics, fault prediction of nodes can be achieved based on the spatial map of the distribution network.

2. The method for fault prediction based on a distribution network spatial map according to claim 1, characterized in that, The process of constructing a spatial graph model of the power distribution network and obtaining spatial feature vectors includes the following steps: A two-dimensional spatial model of the power distribution network is constructed, and a basic spatial map of the power distribution network is generated by overlaying some content of the multi-source data. Based on the aforementioned two-dimensional spatial model of the power distribution network, a spatial diagram model of the power distribution network is constructed according to structural data; the structural data refers to data associated with power grid lines. GCN is used to learn the spatial dependencies of nodes in the spatial graph model of the distribution network, and the local and global spatial feature vectors of the nodes are extracted as spatial dependencies as spatial feature vectors.

3. The method for fault prediction based on distribution network spatial diagrams according to claim 1, characterized in that, The node characteristics include the device's own attributes, the environmental parameters of the area where the node is located, and the historical fault attributes corresponding to the node. The spatial features include topological association features, spatial proximity influence features, and regional common features; wherein, the topological association features include the connection strength between nodes and adjacent devices, and the fault risk transmission coefficient of adjacent nodes; the spatial proximity influence features include the device status of surrounding nodes and the spatial consistency of the regional environment; the regional common features include the average degree of device aging within the same topological partition, the spatial coverage of regional meteorological disasters, and fault susceptibility clustering caused by topography.

4. The method for fault prediction based on distribution network spatial diagrams according to claim 1, characterized in that, The generation of the distribution network fault spatial distribution map refers to: integrating the spatial risk level with the distribution network spatial map, using different colors to mark the regional risk, and generating the distribution network fault spatial distribution map. It receives real-time sensor monitoring data, meteorological and geographical data, and updates the spatial risk level and distribution network fault spatial distribution map according to a certain period.

5. The method for fault prediction based on a distribution network spatial map according to claim 1, characterized in that, The fault prediction of the node refers to the output of the probability of fault occurrence of each node / line in the next 1 to 24 hours through the GCN-BiLSTM ensemble model.

6. The method for fault prediction based on a distribution network spatial map according to claim 5, characterized in that, Constructing the GCN-BiLSTM ensemble model includes the following steps: Obtain the spatial feature vectors and multidimensional time-series data of all nodes; The timestamps and corresponding meteorological data are extracted from multidimensional time-series data as feature sets, and the timestamps and corresponding historical fault attributes are extracted as label sets. The feature sets and label sets constitute the fault prediction dataset. Define the GCN-BiLSTM ensemble model, which includes an input layer, a GCN feature enhancement layer, a BiLSTM temporal modeling layer, and an output layer. The fault prediction dataset is divided into a training set, a validation set, and a test set. The training set is used for parameter learning of the GCN-BiLSTM ensemble model, the validation set is used for hyperparameter tuning of the GCN-BiLSTM ensemble model, and the test set is used for performance evaluation of the GCN-BiLSTM ensemble model. Model evaluation and optimization were performed to complete the construction of the GCN-BiLSTM ensemble model.

7. The method for fault prediction based on a distribution network spatial map according to claim 6, characterized in that, The input layer concatenates spatial feature vectors and multidimensional temporal data to form a "space-time" joint input matrix. The dimensions of the "space-time" joint input matrix include: time step T, number of nodes N, and number of features F, where T=24 and F≥12; the features corresponding to the number of features F cover node features and spatial features.

8. The method for fault prediction based on distribution network spatial diagrams according to claim 1, characterized in that, The preprocessing includes outlier detection and removal, missing value imputation, and data normalization. The outlier detection and removal refers to setting an outlier threshold based on the characteristics of the distribution network data, and identifying and removing sensor data and logically contradictory topology data that exceed the outlier threshold; the missing value filling refers to using a graph attention-based interpolation method to fill in missing spatiotemporal data by utilizing the spatial correlation of adjacent nodes and historical time series data; the data normalization refers to using Z-Score normalization to normalize numerical data and map the data to the [-1,1] interval.

9. The method for fault prediction based on a distribution network spatial map according to claim 1, characterized in that, When performing data format conversion, multi-source data is classified, with data associated with power grid lines being treated as graph structure data and data associated with timestamps being treated as time series data.

Images