A power distribution network-oriented space-time fusion neural network line topology analysis method

Abstract

The application relates to the technical field of model analysis, in particular to a time-space fusion neural network line topology analysis method for a power distribution network. The method comprises the following steps: obtaining original line topology data corresponding to the power distribution network, carrying out structural decomposition and preprocessing, so as to construct a time-space double graph structure; constructing a bidirectional dynamic feature interaction mechanism based on the time-space double graph structure, carrying out multi-scale topology feature extraction, and generating a time-space separated feature vector set; deeply coupling and fusing the time-space separated feature vector set, generating a unified topology feature representation containing an abnormal topology; constructing a dynamic topology state prediction model based on the unified topology feature representation, optimizing a time-space joint loss function, and outputting corresponding real-time topology connection relationships and device state change trends, simultaneously carrying out dynamic topology reconstruction, so as to generate a current time reliable topology graph containing potential branch disconnection and temporary tripping. The application can improve the topology analysis accuracy of the power distribution network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model analysis, in particular to a power distribution network-oriented spatio-temporal fusion neural network line topology analysis method. BACKGROUND

[0002] With the development of modern power systems, the penetration rate of new energy in distribution networks is continuously increasing, which poses new challenges to the dispatching operation, power quality and system protection of distribution networks. On the one hand, the uncertainty in the distribution system has significantly increased, and the operation mode has become more complex and variable. In order to ensure the economy and reliability of the system, network topology reconfiguration needs to be frequently performed, resulting in complex and frequently changing topology structure of the distribution network. On the other hand, distribution network topology identification is closely related to power flow calculation, state estimation, fault location, etc. Accurate topology identification is crucial for the optimal operation of distribution networks with high penetration of new energy.

[0003] At present, the methods for distribution network topology identification mainly include traditional identification methods and artificial intelligence identification methods. Traditional identification methods such as correlation judgment method, signal injection method and linear programming method require complex calculations, which affects the timeliness of online topology identification. Although the topology identification method based on machine learning has been gradually applied, it still has some shortcomings. For example, the common method based on graph convolution network (GCN) limits the input node information to node voltage data, making it difficult to learn deep features of measurement data and unable to distinguish the importance difference of different measurement data, resulting in low accuracy of topology identification. SUMMARY

[0004] Therefore, it is necessary to provide a power distribution network-oriented spatio-temporal fusion neural network line topology analysis method and system to solve at least one of the above technical problems.

[0005] To achieve the above-mentioned purpose, a power distribution network-oriented spatio-temporal fusion neural network line topology analysis method comprises the following steps:

[0006] Step S1: Obtain the original line topology data corresponding to the distribution network, including the account information corresponding to the line operation, the electrical quantity time series data and the device spatial distribution data, and perform structured disassembly and preprocessing on the original line topology data through a spatio-temporal separation encoder to construct a spatio-temporal bipartite graph structure containing a time series tensor and a spatial adjacency matrix, wherein the time series tensor records the dynamic evolution characteristics of the voltage and current corresponding electrical quantities, and the spatial adjacency matrix represents the physical connection relationship of the devices and the initial topology correlation strength;

[0007] Step S2: a bidirectional dynamic feature interaction mechanism is constructed based on a space-time two-dimensional graph structure, so as to model long-term dependence of a time sequence tensor through a bidirectional gate graph recurrent unit in a time dimension, capture bidirectional evolution rules of electrical quantities at multiple time points, extract multi-scale topological features of a spatial adjacency matrix through a hierarchical graph attention pooling module in a space dimension, establish dynamic correlation weights between nodes, generate a set of space-time separated feature vectors, including a time dimension feature vector and a space dimension feature vector;

[0008] Step S3: a cross-modal gate cycle and a graph attention hybrid mechanism are used to deeply couple and fuse the set of space-time separated feature vectors, and a dynamic weighted interaction of the time dimension feature vector and the space dimension feature vector is realized through a learnable gate weight parameter, so as to generate a unified topological feature representation containing an abnormal topological corresponding;

[0009] Step S4: a dynamic topological state prediction model is constructed based on the unified topological feature representation, so as to optimize a space-time joint loss function through end-to-end training, output real-time topological connection relationships and device state change trends of a power distribution network, and dynamically reconstruct a topological graph based on the real-time topological connection relationships and the device state change trends of the power distribution network, so as to generate a reliable topological graph at a current time point containing potential branch disconnection and temporary tripping.

[0010] The beneficial effects of the application are as follows:

[0011] Compared with the prior art, the power distribution network-oriented space-time fusion neural network line topology analysis method has the beneficial effects that: through multi-source data integration and space-time double graph structure construction, the problem of "single data source and incomplete feature representation" of the traditional method is solved from the data level, providing high-quality and structured input basis for subsequent topology identification. The traditional topology identification method often relies on single electrical quantity data (such as only voltage data), which leads to insufficient feature dimension and difficulty in capturing topology dynamic changes. The step first acquires multi-source raw data such as account information, electrical quantity time series data and device spatial distribution data, covering the "static attributes (device connection relationship)" and "dynamic characteristics (electrical quantity evolution)" of the power distribution network topology, avoiding identification deviation caused by one-sided data. Through the space-time separation encoder, the original data is structured and disassembled, the dynamic electrical quantity data is converted into a time series tensor (records the time series change of voltage and current, such as the current trend under load fluctuation), and the static device connection data is converted into a spatial adjacency matrix (characterizes the physical connection and correlation strength of the device, such as the connection weight of the transformer and the line), and the constructed space-time double graph structure not only retains the dynamic evolution law of the electrical quantity, but also clearly defines the spatial topology correlation of the device, breaking through the limitation of the traditional GCN method that only uses node voltage data and has limited feature dimension, thereby laying a foundation for the separate modeling and deep fusion of space-time features. Secondly, through the bidirectional dynamic feature interaction mechanism, the features are deeply mined from the time and space dimensions respectively, solving the problem of "insufficient long-term dependence capture and rough topology feature extraction" of the traditional method, greatly improving the comprehensiveness and accuracy of feature learning. The traditional GCN method cannot effectively process the time series correlation of electrical quantity, and lacks hierarchical sense in extracting spatial topology features, making it difficult to capture topology dynamic changes. In the time dimension, the bidirectional gated graph recurrent unit is adopted, which can model the evolution trend of electrical quantity forward and trace back to the history dependence, effectively capturing the bidirectional evolution law of electrical quantity at multiple time points, avoiding the loss of long-term dependence caused by single direction modeling. In the spatial dimension, the hierarchical graph attention pooling module extracts topology features through multi-scale (such as first extracting the local connection features of a single line, and then integrating the global topology features of the regional power grid), and assigns dynamic correlation weights to different device nodes (such as assigning high weights to key transformer nodes and low weights to secondary branch nodes), solving the problem of "unable to distinguish the importance of measured data" of the traditional method, making the model more focused on key topology correlation and avoiding identification errors caused by incomplete feature extraction.Then, the spatio-temporal feature deep coupling fusion is realized through the cross-modal mixing mechanism, the problem of "spatio-temporal feature segmentation and shallow fusion" of the traditional method is solved, a more recognizable unified topological feature representation is generated, and the identification ability of abnormal topology is improved. The traditional topological identification method often processes time and space features separately or only performs simple splicing, resulting in insufficient feature correlation and difficulty in capturing "spatio-temporal linkage topology anomaly"; in this step, the cross-modal gating cycle and graph attention mixing mechanism are adopted, the fusion ratio of time and space features is dynamically adjusted through the learnable gating weight parameter, and this dynamic weighted interaction not only realizes the deep coupling of spatio-temporal features, but also highlights the feature difference of abnormal topology, thereby providing high-quality feature support for subsequent topology state prediction. Finally, through the dynamic topology state prediction model and real-time reconstruction, the problem of "poor timeliness of online identification and delayed capture of topology anomaly" of the traditional method is solved, and the precise and real-time monitoring of the distribution network topology is realized. Traditional methods such as linear programming are complex and difficult to meet the timeliness requirements of online identification; in this step, an end-to-end prediction model is constructed based on the unified topological feature representation, the model training and inference efficiency is greatly improved through the optimization of the spatio-temporal joint loss function, and the real-time topology connection relationship and device state change trend can be quickly output, meeting the timeliness requirements of online identification. At the same time, based on the prediction results, dynamic topology reconstruction can accurately identify potential branch disconnection, temporary tripping and other problems, and generate a reliable topology graph at the current time. Compared with the traditional method which can only output static topology, the topology state can be updated in real time through dynamic reconstruction, avoiding scheduling decision errors caused by topology information lag, and improving the identification accuracy of topology anomaly, thereby providing key support for the safe operation and precise scheduling of the distribution network. BRIEF DESCRIPTION OF DRAWINGS

[0012] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, when read in conjunction with the accompanying drawings:

[0013] Figure 1 The step flowchart of the spatio-temporal fusion neural network line topology analysis method for the distribution network of the present application is shown in the figure.

[0014] Figure 2 For Figure 1 The detailed step flowchart of step S1 in the figure. DETAILED DESCRIPTION

[0015] The technical method of the present application will be described in detail below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0016] Moreover, the attached drawings are only schematic and are non-limiting. Identical components have been given the same reference numerals in the various drawings and the same reference designators have been maintained where possible in the cross-referenced drawings. DETAILED DESCRIPTION

[0017] It should be understood that, although terms such as "first" and "second" can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to differentiate one element from another. For example, without departing from the scope of the example embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein encompasses any and all combinations of one or more of the associated listed items.

[0018] To achieve the above object, there is provided Figures 1 to 2 The application provides a power distribution network-oriented space-time fusion neural network line topology analysis method, which comprises the following steps:

[0019] Step S1: Obtain the original line topology data corresponding to the power distribution network, including the account information corresponding to the line operation, the electrical quantity time series data and the device spatial distribution data, and perform structured disassembly and preprocessing on the original line topology data through a space-time separation encoder to construct a space-time double graph structure comprising a time series tensor and a spatial adjacency matrix, wherein the time series tensor records the dynamic evolution characteristics of the voltage and current corresponding electrical quantities, and the spatial adjacency matrix represents the physical connection relationship of the devices and the initial topological correlation strength;

[0020] Step S2: Construct a bidirectional dynamic feature interaction mechanism based on the space-time double graph structure, to model the long-term dependence of the time series tensor through a bidirectional gate graph recurrent unit in the time dimension, capture the bidirectional evolution law of the electrical quantity at multiple time points, and at the same time, extract the multi-scale topological features of the spatial adjacency matrix through a hierarchical graph attention pooling module in the spatial dimension, to establish the dynamic correlation weight between nodes, generate a set of space-time separated feature vectors, including time dimension feature vectors and spatial dimension feature vectors;

[0021] Step S3: Deeply couple and fuse the set of space-time separated feature vectors by using a cross-modal gate cycle and graph attention hybrid mechanism, and realize the dynamic weighted interaction of the time dimension feature vectors and the spatial dimension feature vectors through learnable gate weight parameters, to generate a unified topological feature representation comprising the abnormal topological.

[0022] Step S4: based on the unified topological feature representation, a dynamic topological state prediction model is constructed to optimize the spatio-temporal joint loss function through end-to-end training, and output the corresponding real-time topological connection relationship and device state change trend of the power distribution network, and simultaneously based on the corresponding real-time topological connection relationship and device state change trend of the power distribution network, the dynamic topological reconstruction is carried out to generate the current time reliable topological graph containing potential branch disconnection and temporary trip.

[0023] In the embodiment of the application, please refer to Figure 1 As shown in the step flow diagram of the power distribution network-oriented spatio-temporal fusion neural network line topology analysis method of the application, in the present example, the power distribution network-oriented spatio-temporal fusion neural network line topology analysis method comprises the following steps:

[0024] Step S1: obtaining the original line topology data corresponding to the power distribution network, including the account information corresponding to the line operation, the electrical quantity time series data and the device spatial distribution data, and through the spatio-temporal separation encoder, the original line topology data is structured and preprocessed to construct the spatio-temporal double graph structure containing the time series tensor and the space adjacency matrix, wherein the time series tensor records the dynamic evolution characteristics of the voltage and current corresponding electrical quantity, and the space adjacency matrix represents the device physical connection relationship and the initial topological correlation strength;

[0025] In the embodiment of the application, by acquiring the original line topology data of the IEEE33 node power distribution network, including line account information (32 line models such as JKLYJ-10-120, connection node numbers such as line 1 connecting nodes 1-2, 20 overhead and 12 cable laying modes), electrical quantity time series data (1 minute sampling interval, 72 hours of 33 node three-phase voltage 0.9-1.1pu, current 0-250A, active power 0-600kW, and reactive power 0-300kVar, generating a 33x4x4320 data set), and device spatial distribution data (substation node 1 latitude and longitude 30.1234°N / 120.5678°E, 32 base towers with latitude and longitude accuracy of 0.0001° and elevation accuracy of 0.1m). Through the spatio-temporal separation encoder preprocessing: missing values are filled with linear interpolation (such as node 5 voltage from 0.98pu to 1.02pu at the 1000th timestamp), misconnections are corrected by the isolation forest algorithm (such as line 8 misconnection nodes 7-9 corrected to 7-8), and electrical quantities are normalized to the 0-1 interval by range. Disassemble to generate time series tensor (33x4x432, take 1 time step every 10 minutes, record electrical quantity dynamic evolution) and spatial adjacency matrix (33x33, element value=1 / (1+line impedance), impedance is calculated according to overhead 0.08Ω / km and cable 0.05Ω / km, such as node 1-2 impedance 0.096Ω corresponding element 0.912), construct a spatio-temporal bipartite graph structure, and ensure that the time tensor index and the matrix node are one-to-one corresponding.

[0026] Step S2: based on the spatio-temporal bipartite graph structure, a bidirectional dynamic feature interaction mechanism is constructed to model the long-term dependence of the time series tensor through the bidirectional gated graph recurrent unit in the time dimension, capture the bidirectional evolution law of the electrical quantity at multiple time points, and at the same time, through the hierarchical graph attention pooling module in the spatial dimension, the multi-scale topological feature of the spatial adjacency matrix is extracted to establish the dynamic correlation weight between nodes, and generate a set of feature vectors separated in space and time, including time dimension feature vectors and spatial dimension feature vectors;

[0027] In the embodiment of the application, a bidirectional dynamic feature interaction mechanism is constructed based on a space-time two-dimensional graph structure. A bidirectional gated graph recurrent unit (Bi-GGRU) is used in the time dimension to model a 33x4x432 time sequence tensor: forward propagation from the 1st to the 432th time step, input gate weights [0.7, 0.65, 0.72, 0.68] are used to filter the voltage power new features, and forgetting gates [0.2, 0.25, 0.18, 0.22] are used to discard redundant information to generate a forward state vector; backward propagation from the 432th to the 1st time step, reference is made to features of the next 5 time steps, backward gate weights [0.72, 0.68, 0.74, 0.70] are used to capture future trends to generate a backward state vector, a 33x432 time dimension feature vector is obtained by fusion, and the long-term dependence information retention rate is greater than or equal to 83%. A hierarchical graph attention pooling (HGAP) is used in the space dimension to extract multi-scale features from a 33x33 adjacency matrix: a bottom attention head is used to calculate the weight of directly adjacent nodes (the similarity between nodes 5-4 is 0.98x0.6+ the distance 0.8km attenuation 0.556x0.4=0.822), 3 hierarchical jumps are used to capture 2-3 order neighborhoods (the weight of the 1st hierarchical 2-order neighborhood node 5-3 is 0.65), and global pooling is used to give regional weights (root region 0.84, branch 0.75, and terminal 0.66) to generate a 33x33 space dimension feature vector, and the key node feature distinction degree is higher than that of a traditional GCN.

[0028] Step S3: a cross-modal gated cycle and a graph attention hybrid mechanism are used to deeply couple and fuse the space-time separated feature vector set, and a learnable gating weight parameter is used to realize dynamic weighted interaction between the time dimension feature vector and the space dimension feature vector to generate a unified topological feature representation corresponding to an abnormal topological structure.

[0029] In the embodiment of the application, the feature vector set is fused by adopting a cross-modal gating cycle and a graph attention hybrid mechanism. The time dimension feature vector (33x432) is flattened to 33x14256, the space dimension (33x33) is flattened to 33x1089, the dimensions are unified to 33x1089 through a full connection network, and the L2 standardization is used to eliminate the dimension. The cross-modal gating unit inputs the spliced 33x2178 vector, a 2-layer full connection network generates a 33x1089 dynamic fusion weight matrix through sigmoid (the node 5 voltage time sequence and the space correlation strong place weight is 0.8, and the weak place is 0.3). The gating cycle transmits the time characteristics: the voltage time sequence rises by 0.2pu to convert into a space weight adjustment signal, and the node 5-4 weight rises from 0.73 to 0.81; the graph attention coefficient feeds back the space characteristics: the node 5-4 attention coefficient 0.82 corrects the time dependent weight from 0.5 to 0.65, and enhances the abnormal identification. The 33x1089 unified topology feature representation is generated, the abnormal topology is like the feature value 0.05-0.072 of the node 8-9 account error connection, which is significantly higher than the normal area 0.002-0.03, and the prediction accuracy of the IEEE 33 node test set is higher than that of simple splicing.

[0030] Step S4: constructing a dynamic topology state prediction model based on the unified topology feature representation, optimizing a space-time joint loss function through end-to-end training, and outputting the corresponding real-time topology connection relationship and device state change trend of the power distribution network, and simultaneously performing dynamic topology reconstruction based on the corresponding real-time topology connection relationship and device state change trend of the power distribution network to generate a current time reliable topology graph containing potential branch disconnection and temporary trip.

[0031] In this embodiment of the invention, a dynamic topology state prediction model is constructed based on a unified topology feature representation. The model includes a 3-layer graph convolutional decoder (layer 1: 3×3 kernel, input 1089, output 512; layer 2: 3×3 kernel, output 256; layer 3: 1×1 kernel, output 33; Softmax generates a 33×33 connection probability matrix). The spatiotemporal joint loss function includes a temporally improved Huber loss (threshold 0.12 pu; small deviation squared loss, such as 0.08 pu corresponding to 0.0032; large deviation absolute value loss, such as 0.2 pu corresponding to 0.0168), a spatial cross-entropy loss (critical path weighting factor 0.8-0.9, non-critical 0.3), and a link prediction loss (Bayes ranking loss optimizing edge weight ranking). The dynamic weighting is 0.5:0.5 for the first 50 rounds and transitions to 0.4:0.6 for rounds 51-150. The AdamW optimizer (initial learning rate 0.005, cosine annealing cycle 50 epochs, early stop 15 epochs) was used for end-to-end training with a batch size of 32, L2=1e-4+Dropout=0.3. After 180 training epochs, the validation set loss stabilized at 0.017, and the test set topology accuracy was 96.2%. Inputting 33×1089 features generated from one hour of real-time data, the model outputs real-time topology connectivity (node ​​1-2 probability 0.99, 8-9 0.26), device health (node ​​8 score 772, 195), and trend (node ​​8-9 has a 78% chance of becoming disconnected in the next hour). Dynamic topology reconstruction marks potentially disconnected nodes (8-9) and normal branches, generating a reliable topology map for the current time, with maintenance priorities and backup paths. Connectivity is 100%, redundancy is 1.8, and load coverage is 100%, meeting safe operation requirements.

[0032] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart of step S1 is shown below. In this embodiment, step S1 includes the following steps:

[0033] Step S11: Collect the original line topology data of the distribution network in collaboration with the distribution network monitoring system, equipment ledger management platform and geographic information system. The ledger information corresponding to the line operation includes the line model, connection node number and laying method information. The electrical quantity time series data includes the effective values ​​of three-phase voltage and current and active and reactive power data sampled at the minute level. The equipment spatial distribution data includes the latitude, longitude and elevation data of substations, switching stations and line towers.

[0034] In the embodiments of the present application, the original line topology data of the IEEE33 node power distribution network is collected by the power distribution network monitoring system, the equipment account management platform and the geographic information system. The power distribution network monitoring system collects the three-phase voltage (range 0.9-1.1pu), current effective value (range 0-250A) and active power (range 0-600kW), reactive power (range 0-300kVar) data of 33 nodes with a sampling interval of 1 minute, continuously collects for 72 hours, and generates an electrical quantity time series data set (33 nodes, 4 types of parameters, 4320 time stamps). The equipment account management platform exports line account information, including the model (such as JKLYJ-10-120) of 32 lines, the connection node number (such as line 1 connecting node 1 and node 2) and the laying method (overhead / cable, of which 20 are overhead and 12 are cable). The geographic information system obtains the longitude and latitude (accuracy 0.0001°) and elevation (accuracy 0.1m) data of the transformer station (node 1), the switch station (node 10, 20) and the 32 base line towers, such as the longitude and latitude of node 1 30.1234°N, 120.5678°E, and the elevation 50.2m, the longitude and latitude of tower 5 30.1245°N, 120.5689°E, and the elevation 52.5m. During the collection process, the system time synchronization tool is used to ensure that the time stamp deviation of the three systems is ≤1 second, the account number and the node number are one-to-one corresponding, and the complete original data covering equipment, electrical quantity and spatial position are generated.

[0035] Step S12: preprocessing the original line topology data to fill in the missing values in the electrical quantity time series data and equipment spatial distribution data using interpolation algorithm, and using isolated forest algorithm to identify and correct the abnormal records corresponding to the line misconnection and missing connection points in the account information, and eliminating the influence of different dimension electrical parameters by range normalization method, to generate line topology standardized data;

[0036] In the embodiments of the present application, the original line topology data is preprocessed. For the 5 missing values in the electrical quantity time series data (such as the missing voltage data of node 5 at the 1000th timestamp), a linear interpolation algorithm is used to fill in: based on the 5 valid data points before and after the missing value (995-999, 1001-1005 timestamps), the missing value is calculated according to the time interval ratio (such as the voltage linearly transitions from 0.98pu to 1.02pu, and the missing value is filled in as 1.00pu); for the missing elevation of 2 base towers in the device spatial distribution data, the adjacent interpolation (taking the average of the elevations of the surrounding 3 base towers) is used to fill in. The isolated forest algorithm (100 trees, abnormal threshold 0.8) is used to identify the abnormal records in the account: the misconnection record of line 8 "connecting node 7 and node 9" (the actual connection is node 7 and node 8), and the "missing connection point" record of line 15 are detected, and combined with the geographic information system tower distribution (the towers of line 8 are distributed along nodes 7-8) and the power flow of the monitoring system (line 15 has power transmission proof that the connection point exists), the correct account is corrected. The range normalization is used for the electrical parameters: the voltage is converted according to (actual value-0.9) / (1.1-0.9), and the current is converted according to (actual value-0) / (250-0), so that all parameters are normalized to the interval of 0-1, and finally the line topology standardized data is generated.

[0037] Step S13: inputting the line topology standardized data into a space-time separation encoder, structurally decomposing according to the time dimension and the space dimension, generating a time series tensor indexed by the timestamp and a space adjacency matrix with the node-edge relationship as the core, wherein the dimension of the time series tensor is determined by the sampling time length and the number of electrical parameter types, and the initial value of the element of the space adjacency matrix is calculated by the physical distance between devices and the line impedance;

[0038] In the embodiments of the present application, the line topology standardization data is input into the space-time separation encoder for structured disassembly. In the time dimension, the normalized voltage, current, active power and reactive power parameters of each node are extracted with 1-minute timestamps as indexes. Considering the subsequent processing efficiency, every 10-minute data is selected as a time step, and 72 hours are divided into 432 time steps, thereby constructing a time series tensor of 33x4x432 (33 nodes, 4 types of parameters, and 432 time steps). The dimensions are determined by the sampling time length (72 hours / 10 minutes=432) and the number of parameter types (4). In the spatial dimension, the node-edge relationship is taken as the core, the physical distance between nodes (for example, the distance between node 1 and node 2 is 1.2 km) is calculated according to the device latitude and longitude, the line impedance (the impedance of line 1 is 1.2x0.08=0.096 Ω) is obtained by using the impedance calculation formula (the impedance of overhead line is 0.08 Ω / km, and the impedance of cable is 0.05 Ω / km), and a 33x33 spatial adjacency matrix is constructed. The initial value of the matrix element is 1 / (1+impedance) (the smaller the impedance, the larger the element value, which represents the connection strength), for example, the element value of node 1-2 is 1 / (1+0.096)=0.912, and the element value of non-connected nodes is 0. After disassembly, the dimensions are checked to ensure that the tensor time steps correspond to the actual sampling, and the adjacency matrix elements are matched with the impedance one by one, thereby generating structured data.

[0039] Step S14: dynamically correcting the spatial adjacency matrix to adjust the weight of the initial value of the element in the spatial adjacency matrix by introducing the real-time operation parameters of the line, including the power flow distribution and the voltage drop rate, so that the spatial adjacency matrix reflects the physical connection relationship and the electrical function coupling strength of the device, and generates a space-time double graph structure in combination with the time series tensor.

[0040] In the embodiment of the present application, the spatial adjacency matrix is dynamically corrected. Real-time operation parameters of the line are introduced: power flow distribution (the normalized power flow of each line is obtained from the monitoring system, ranging from 0 to 1), voltage drop rate (the voltage difference between adjacent nodes is calculated / distance, ranging from 0 to 0.1). The weight adjustment formula is set: corrected element value = initial value x (0.6 x power flow + 0.4 x (1-voltage drop rate)), for example, line 1 initial value 0.912, power flow 0.8, voltage drop rate 0.05, corrected = 0.912 x (0.6 x 0.8 + 0.4 x 0.95) = 0.912 x (0.48 + 0.38) = 0.912 x 0.86 = 0.784; line 5 initial value 0.85, power flow 0.3, voltage drop rate 0.1, corrected = 0.85 x (0.18 + 0.36) = 0.85 x 0.54 = 0.459. The corrected matrix element value reflects both the physical connection (initial value) and the electrical function (high power flow and small voltage drop, the value is large). The corrected spatial adjacency matrix is combined with the 33x4x432 time series tensor to construct a space-time double graph structure: the nodes of the graph are the nodes of the distribution network, the edge weight is the corrected value of the adjacency matrix, and the node attribute is the corresponding parameter of the time series tensor, realizing the fusion of physical connection and electrical time sequence characteristics, and providing complete graph structure data for subsequent neural network analysis.

[0041] Further, the space-time separation encoder in step S13 comprises a feature mapping layer, a dimension alignment module and a space-time consistency verification unit, and the specific implementation process comprises:

[0042] Step S131: The feature mapping layer structures and disassembles the line topology standardized data according to the time dimension and the space dimension through a multi-modal embedding function, to generate a time series tensor indexed by a time stamp and a spatial adjacency matrix with a node-edge relationship as the core;

[0043] In the embodiment of the application, the feature mapping layer adopts a space-time separation encoder to structurally disassemble the line topology standardized data of the IEEE33 node power distribution network (including voltage, current, power and other electrical parameters of 33 nodes, a sampling frequency of 5 minutes / time, a total of 72 hours of data collected, and known node spatial coordinates). The multi-modal embedding function is split according to the time dimension and the space dimension: in the time dimension, with a 5-minute timestamp as the index, the voltage (range 0.95-1.05pu), current (range 0-200A) and power (range 0-500kW) parameters of each node are extracted to construct a 33x3x288 time series tensor (33 nodes, 3 types of electrical parameters, 288 timestamps); in the space dimension, taking the node-edge relationship as the core, the physical distance (unit: km) is calculated according to the node spatial coordinates, and a distance ≤1km is set as a connected edge to construct a 33x33 space adjacency matrix, with the matrix element 1 representing connectivity and 0 representing non-connectivity, and the line impedance unit value (range 0.092-0.5) is supplemented as the edge weight to the adjacency matrix, so that the matrix reflects both physical connectivity and impedance correlation. In the disassembly process, dimension checking is performed to ensure that the time series tensor index and the timestamp are one-to-one corresponding, the node order of the space adjacency matrix is consistent with the actual power grid node number, and the generated tensor and matrix provide a structured data basis for subsequent space-time processing.

[0044] Step S132: The dimension alignment module synchronizes the time axes of data with different sampling frequencies through a dynamic time warping algorithm, processes the time series tensor with high-frequency sampling by sliding window mean down-sampling, and processes the space adjacency matrix updated at low frequency by linear interpolation up-sampling, so that the time series tensor and the space adjacency matrix are consistent in the time scale, avoiding feature misalignment caused by differences in data sampling frequency;

[0045] In the embodiment of the application, the dimension alignment module synchronizes the time axis by using the dynamic time warping algorithm for the sampling frequency difference between the time series tensor (sampling frequency 5 minutes / time) and the spatial adjacency matrix (original update frequency 2 hours / time). For the high-frequency time series tensor, a sliding window mean down-sampling processing is used: the window size is set to 24 (corresponding to 2 hours), the mean value of the electrical parameters (such as the voltage mean value 1.02pu and the current mean value 150A) in each window is calculated, the 288 time stamps are down-sampled to 12, and a down-sampled time series tensor of 33x3x12 is generated; for the low-frequency spatial adjacency matrix, a linear interpolation up-sampling processing is used: based on the original adjacency matrix of 2 hours interval (a total of 36), 23 interpolation matrices are inserted between the adjacent two matrices, the edge weight at the intermediate time (such as the impedance of 0.2 at t1 and 0.25 at t2, and the interpolation is calculated according to 0.2+0.05x(t-t1) / (t2-t1) at the intermediate time) is calculated by linear calculation, the 36 matrices are up-sampled to 288, and are further down-sampled to 12, which is consistent with the time scale of the time series tensor. After synchronization, the time stamp matching check is performed to ensure that both correspond to 12 same time nodes (one every 2 hours), so as to avoid the problem that the voltage fluctuation and the topology change are not synchronized due to the frequency difference.

[0046] Step S133: The space-time consistency verification unit identifies the asynchronous deviation of the time series data and the spatial data by calculating the matching degree of the mutation time corresponding to the time series tensor and the time stamp of each topology structure change record in the spatial adjacency matrix, and starts a secondary verification mechanism for the samples with asynchronous deviation exceeding a set threshold, so as to correct the asynchronous deviation by combining the historical topology evolution rule and the electrical quantity change correlation rule, and generate the processed time series tensor and the spatial adjacency matrix.

[0047] In the embodiment of the application, the spatio-temporal consistency verification unit calculates the matching degree of the mutation time of the time series tensor and the timestamp of the topology change record of the spatial adjacency matrix. First, the mutation time in the time series tensor is identified by a mutation detection tool (the threshold is set to the mean value of the electrical parameter ± 15%): for example, the voltage of node 5 decreases from 1.03 pu to 0.96 pu at the 6th timestamp (10 hours), which is determined as a mutation time; the topology change record in the spatial adjacency matrix is extracted (for example, the edge weight of node 5 and node 6 changes from 0.22 to 0.35 at the 5th timestamp), and the time difference between the mutation time and the change timestamp (2 hours in this case) is calculated. The matching degree threshold is set to 1 hour (the time difference ≤ 1 hour is matched), and the sample time difference of 2 hours exceeds the threshold, so the secondary verification mechanism is started. Combined with the historical topology evolution rule (the topology change of node 5 in the past 3 months is more accompanied by voltage fluctuation) and the electrical quantity change correlation rule (impedance increase of 15% corresponds to voltage drop of 5%-8%), the change timestamp of the spatial adjacency matrix is corrected to the 6th timestamp, so that the change is synchronized with the voltage mutation, and the processed time series tensor and the spatial adjacency matrix with a time difference of 0 hours are generated. After verification, the compliance rate of all samples is ≥ 95%, ensuring the spatio-temporal data consistency.

[0048] Step S134: Dimension compression and feature selection are performed on the processed time series tensor and the spatial adjacency matrix, so as to extract key time sequence features in the time series tensor by using principal component analysis technology, and simplify the spatial adjacency matrix by using a sparse matrix decomposition method, while retaining core information and reducing data calculation complexity.

[0049] In the embodiment of the application, dimension compression and feature selection are performed on the processed 33x3x12 time series tensor and the 33x33 spatial adjacency matrix. Key time sequence features are extracted by using principal component analysis technology: the covariance matrix (99x99, 33 nodesx3 parameters) of the tensor is calculated, the eigenvalues are solved and sorted according to the proportion, the first two principal components with a cumulative contribution rate ≥ 90% are selected, a 99x2 dimension reduction matrix is constructed, the original tensor is converted into a 33x2x12 dimension reduction time sequence tensor, and the core change trend of voltage and current is retained; the spatial adjacency matrix is simplified by using a sparse matrix decomposition method: the 33x33 matrix is decomposed into a lower triangular matrix and an upper triangular matrix by using LU decomposition, the rows and columns with a non-zero element proportion ≥ 10% (corresponding to the core connected branch of the power grid) are retained, isolated nodes and invalid edges are removed, a 28x28 sparse adjacency matrix is generated, and the non-zero elements still retain the original edge weight. After compression, the data calculation amount is reduced by 40%, and at the same time, the information retention rate is verified: the time sequence feature retains 92% of the original information, and the adjacency matrix connectivity description accuracy is 98%, which reduces the complexity while ensuring the integrity of the core topology and electrical quantity information.

[0050] Further, the bidirectional gated graph recurrent unit in step S2 comprises a forward propagation chain, a backward propagation chain, a state fusion gate and a residual connection module, and the long-term dependence modeling of the time series tensor by the bidirectional gated graph recurrent unit in the time dimension comprises:

[0051] The forward propagation chain encodes the time series tensor from the historical time to the current time to filter the new features of the current time through the input gate, discard the redundant historical information through the forget gate, and generate the forward state vector of the current time through the output gate, wherein the gating weight parameters are determined by the electrical quantity features of the current time and the historical state features;

[0052] In the embodiment of the present application, the forward propagation chain encodes the 33x4x432 time series tensor (4 types of parameters of node 5: voltage V, current I, active power P, and reactive power Q, time steps t=1 to t=432) from t=1 to t=432. The forward gating structure of the bidirectional gated graph recurrent unit is adopted, and the input gate, the forget gate, and the output gate are all 3-layer fully connected networks (input dimension 4, hidden layer dimension 8, and output dimension 4). For the electrical quantity features of node 5 at t=100 (V=0.85, I=0.6, P=0.7, and Q=0.5, all normalized values), the input gate calculates the weight parameters: combining the current features and the historical state features at t=99 (V=0.83, I=0.58, P=0.68, and Q=0.48), the input gate weight [0.7, 0.65, 0.72, 0.68] is output through the Sigmoid activation function, which filters the significant changes of voltage and power in the current new features; the forget gate weight [0.2, 0.25, 0.18, 0.22] discards the redundant historical information of current and reactive power; and the output gate weight [0.8, 0.75, 0.82, 0.78] combines the filtered current features with the retained historical features to generate the forward state vector at t=100 [0.84, 0.59, 0.70, 0.49] through the Tanh activation function. All time steps are calculated according to this logic, the gating weight parameters are iteratively optimized through back propagation, and the forward state vector can accumulate effective historical time series information.

[0053] Preferably, the backward propagation chain decodes the time series tensor from the future time to the current time to generate the backward state vector of the current time by using the gating structure symmetrical to the forward propagation chain, wherein the gating weight parameters are determined by learning the future electrical quantity change trend;

[0054] In the embodiment of the application, the backward propagation chain decodes the time series tensor from t=432 to t=1 direction feature, adopts a symmetrical gating structure (same as the 3-layer fully connected network, the input and output dimensions are consistent) with the forward propagation chain, and the gating weight parameters are determined by learning the future trend of electrical quantity change. For node 5 at t=100, the backward propagation takes the future features (such as V=0.86, I=0.61, P=0.71, Q=0.50 at t=101 to t=105) as reference, calculates the reverse input gate weight [0.72, 0.68, 0.74, 0.70], and focuses on the future upward trend of voltage and power; the reverse forgetting gate weight [0.18, 0.22, 0.16, 0.20] discards the historical fluctuations irrelevant to the future trend; the reverse output gate weight [0.82, 0.78, 0.84, 0.80] fuses the future trend information and the current feature to generate the reverse state vector [0.85, 0.60, 0.71, 0.50] at t=100. The reverse gating weight is trained by minimizing the error between the future feature and the current decoding feature, to ensure that the reverse state vector can capture the future evolution law of the electrical quantity in advance.

[0055] Preferably, the state fusion gate dynamically weights and fuses the forward state vector and the reverse state vector through an attention mechanism, calculates the similarity of the forward state vector and the reverse state vector in different feature dimensions, assigns fusion weights according to the similarity, and generates a comprehensive state vector containing bidirectional time series information;

[0056] In the embodiment of the application, the state fusion gate dynamically weights and fuses the forward state vector (at time t=100, [0.84, 0.59, 0.70, 0.49]) and the reverse state vector ([0.85, 0.60, 0.71, 0.50]) through an attention mechanism. The cosine similarity of the two in each feature dimension is calculated: voltage dimension similarity = 0.84*0.85 / (sqrt(0.84^2+0.59^2+0.70^2+0.49^2)*sqrt(0.85^2+0.60^2+0.71^2+0.50^2))≈0.998, current dimension≈0.997, power dimension≈0.998, and reactive power dimension≈0.997. According to the similarity, the fusion weight is allocated: the higher the similarity, the greater the weight, and the voltage weight is 0.502, the current weight is 0.501, the power weight is 0.502, and the reactive power weight is 0.501. The comprehensive state vector is obtained by weighting according to the weight: V=0.84*0.502+0.85*0.498≈0.845, I=0.59*0.501+0.60*0.499≈0.595, P=0.70*0.502+0.71*0.498≈0.705, and Q=0.49*0.501+0.50*0.499≈0.495, i.e., [0.845, 0.595, 0.705, 0.495], which contains bidirectional time series information of both historical accumulation and future trend.

[0057] Preferably, the residual connection module element-wise adds the original time series tensor and the comprehensive state vector, and alleviates the gradient vanishing problem through layer normalization processing, while introducing an attention mechanism to give different weights to time series features of different electrical parameter types, so as to capture the bidirectional evolution law of electrical quantities at multiple time points to generate a time dimension feature vector.

[0058] In the embodiment of the application, the residual connection module element-wise adds the feature of node 5 at t=100 in the original time series tensor ([0.85, 0.6, 0.7, 0.5]) and the integrated state vector ([0.845, 0.595, 0.705, 0.495]) to obtain the residual feature [1.695, 1.195, 1.405, 0.995]. Through layer normalization processing (calculate the mean 1.3225 and the standard deviation 0.275 of the residual feature, and convert according to (element value-mean) / standard deviation), the normalized residual feature [1.35, -0.46, 0.30, -1.20] is obtained, which alleviates the gradient disappearance problem. The parameterized attention mechanism is introduced to give differentiated weights to the four types of electrical parameters: based on the characteristics that voltage and power are more sensitive to topological changes in power distribution network topology analysis, the voltage weight 0.35, the power 0.35, the current 0.2, and the reactive power 0.1 are determined through training, and the time dimension feature vector is generated after weighting: 1.35x0.35+(-0.46)x0.2+0.30x0.35+(-1.20)x0.1≈0.473-0.092+0.105-0.12=0.366. All 33 nodes and 432 time steps are processed in this way to generate a 33x432 time dimension feature vector matrix, which completely captures the bidirectional evolution rule of electrical quantities.

[0059] Further, the hierarchical graph attention pooling module in step S2 includes multiple hierarchical attention heads, feature enhancement units, and topology robustness correctors. The multi-scale topology feature extraction of the spatial adjacency matrix by the hierarchical graph attention pooling module in the spatial dimension includes:

[0060] The bottom attention head focuses on the local topology features of the directly adjacent nodes, calculates the electrical quantity similarity and the physical distance attenuation coefficient, and fuses the two as the bottom attention weight;

[0061] In the embodiment of the application, the bottom attention head is aimed at a 33x33 spatial adjacency matrix (the direct neighboring nodes of node 5 are 4, 6 and 10), and the local topological feature of the direct neighboring nodes is calculated to obtain the weight. The electrical quantity similarity tool is used to calculate the voltage and power similarity between node 5 and the neighboring nodes: the voltage of node 5 is 0.845, the power is 0.705, the voltage of node 4 is 0.83, the power is 0.69, the similarity = (0.845x0.83+0.705x0.69) / (sqrt(0.845^2+0.705^2)x sqrt(0.83^2+0.69^2))≈0.98; the similarity between node 5 and node 6 is 0.97, and the similarity between node 5 and node 10 is 0.95. The physical distance attenuation coefficient tool is used to calculate the attenuation coefficient = 1 / (1+distance) according to the distance (the distance between node 5 and node 4 is 0.8km, the distance between node 5 and node 6 is 1.0km, and the distance between node 5 and node 10 is 1.5km), and the attenuation coefficients are 0.556, 0.5 and 0.4 respectively. The similarity and the attenuation coefficient are fused by weighting 0.6 and 0.4: the weight between node 5 and node 4 = 0.98x0.6+0.556x0.4≈0.6+0.222=0.822; 5-6≈0.97x0.6+0.5x0.4=0.582+0.2=0.782; 5-10≈0.95x0.6+0.4x0.4=0.57+0.16=0.73. The weight of the non-direct neighboring nodes is set to 0, and a 33x33 bottom attention weight matrix is generated to ensure that the weight focuses on the local correlation of the direct neighboring nodes.

[0062] Preferably, the middle attention head captures the intermediate scale topological features of the 2-N order neighborhood through the jump connection, the receptive field range of each level attention head is expanded exponentially, and the cross-layer residual connection is introduced to prevent feature degradation, and the middle attention weight is calculated by the local topological features and the global topological structure features.

[0063] In the embodiment of the application, the middle layer attention head is provided with 3 levels (N=3), and the 2-3 order neighborhood features are captured through the skip connection. The first level (2-order neighborhood): the 2-order neighborhood of node 5 is nodes 3, 7 and 9 (extended through nodes 4 and 6), the receptive field range is expanded to 2 hops, the similarity of the local features (voltage deviation ≤0.02) and the global features (regional power flow ≥0.6) between the neighborhood nodes and node 5 is calculated, the weight = local similarity × 0.5 + global similarity × 0.5, the weight of node 5-3 is 0.65, the weight of node 5-7 is 0.62, and the weight of node 5-9 is 0.58; the second level (3-order neighborhood): the receptive field is expanded exponentially (covering 8 nodes), the weight calculation introduces a cross-layer residual connection, and 30% of the first level weight is added to the current level, the weight of node 5-2 is 0.55 (including 0.65×0.3 residual); the third level (N=3): the receptive field covers 15 nodes, and the weight is fused with the residual of the previous two levels and the current features, the weight of node 5-1 is 0.52. All the output features of the middle layer attention head are normalized (mean 0.58, standard deviation 0.12) to prevent degradation, and a 33×33 middle layer attention weight matrix is generated, and the receptive field range of each level is expanded by 2¹, 2² and 2³ nodes to ensure complete capture of the intermediate scale topological features.

[0064] Preferably, the high layer attention head extracts the global topological features of the power distribution network through a global pooling operation, and combines the tree structure characteristics of the power grid topology and the power flow distribution law to give different weights to the topological features of different regions.

[0065] In the embodiment of the application, the high layer attention head performs a global pooling operation on the 33-node power distribution network (tree structure, root node 1 is a substation) to extract global topological features. The global average pooling tool is used to calculate the average power flow (root area 0.75, branch area 0.6, and terminal area 0.45) and voltage stability (root area 0.9, branch area 0.85, and terminal area 0.8) of each region (root node area 1-5, branch area 6-15, and terminal area 16-33). Combined with the tree structure characteristics (root node affects the global) and the power flow law (power is concentrated in the root area), different weights are given: root area weight = 0.75×0.4 + 0.9×0.6 = 0.3 + 0.54 = 0.84; branch area = 0.6×0.4 + 0.85×0.6 = 0.24 + 0.51 = 0.75; terminal area = 0.45×0.4 + 0.8×0.6 = 0.18 + 0.48 = 0.66. The regional weight is distributed to the corresponding nodes, node 1 weight 0.84, node 6-15 weight 0.75, and node 16-33 weight 0.66, to generate a global attention weight vector, ensuring that the global features focus on key areas.

[0066] Preferably, the feature enhancement unit performs channel attention weighting on the feature channels output by each hierarchical attention head, calculates the contribution of different feature channels to topology state recognition, and weights and corrects the topology feature information corresponding to each hierarchical attention head based on the contribution of different feature channels to topology state recognition.

[0067] In the embodiment of the present application, the feature enhancement unit calculates the channel contribution degree for the three channels of bottom layer (33x33), middle layer (33x33) and high layer (33x1) attention features. The channel attention tool is used to learn the contribution degree through a fully connected network (input 3D, output 3D): the bottom layer feature captures local connection, the contribution degree is 0.4; the middle layer feature captures intermediate association, the contribution degree is 0.35; the high layer feature captures global structure, the contribution degree is 0.25. The hierarchical features are weighted and corrected according to the contribution degree: bottom layer feature x 0.4 + middle layer feature x 0.35 + high layer feature x 0.25, such as the corrected feature of node 5 = 0.822x0.4 (bottom layer 5-4 weight) + 0.65x0.35 (middle layer 5-3 weight) + 0.84x0.25 (high layer node 5 root area weight) ≈ 0.329 + 0.228 + 0.21 = 0.767. All nodes are corrected in this way to generate a 33x33 enhanced spatial feature matrix, and the channel contribution degree is verified by training data (the accuracy of topology recognition after correction is improved by 5%), which ensures that the feature highlights the key channel information.

[0068] Preferably, the topology robust corrector calculates the feature consistency score to dynamically adjust the weight proportion corresponding to the abnormal connection when it is identified that the abnormal connection contains a misconnection or a short connection of the account, and automatically reduces the weight of the line in the spatial adjacency matrix when it is detected that the topology feature of the line is seriously inconsistent with the surrounding lines, so as to establish the dynamic association weight between nodes and generate a spatial dimension feature vector.

[0069] In the embodiment of the application, the topology robust modifier detects the account error record in the spatial adjacency matrix (node 8 marks the connection between nodes 7 and 9, and actually connects 7 and 8), calculates the feature consistency score: the voltage similarity of node 8 and node 9 is 0.8 (lower than the normal threshold 0.9), the power flow deviation is 0.2 (higher than the threshold 0.1), and the consistency score = 0.8 x 0.5-0.2 x 0.5 = 0.3 (lower than the qualified threshold 0.7). Dynamically adjust the abnormal connection weight: the original weight of node 8-9 is 0.7, and after correction = 0.7 x (consistency score / qualified threshold) = 0.7 x (0.3 / 0.7) = 0.3; At the same time, supplement the correct connection of node 8-8 (the original weight is 0), calculate the similarity 0.96, the attenuation coefficient 0.5, and the weight = 0.96 x 0.6 + 0.5 x 0.4 = 0.776. For the short circuit anomaly (nodes 12 and 13 are short-circuited, and the power flow suddenly increases), the consistency score is 0.2, and the short-circuit connection weight is reduced from 0.8 to 0.23. After correction, a 33x33 dynamic correlation weight matrix is generated, the matrix is flattened into a 33x33 spatial dimension feature vector according to the row, and it is ensured that the vector accurately reflects the corrected topology correlation.

[0070] Further, step S3 comprises the following steps:

[0071] Step S31: Extract the generated time dimension feature vector and spatial dimension feature vector, map the two types of features to the same dimension space through the feature standardization layer, and adopt the L2 standardization method to eliminate the feature dimension difference, so as to ensure that the time-space features are fused in the same scale;

[0072] In the embodiment of the application, the 33x432 time dimension feature vector matrix (the feature value range of node 5 at 432 time steps is 0.3-0.9) and the 33x33 spatial dimension feature vector matrix (the correlation weight range of node 5 and other nodes is 0.2-0.9) are extracted, and are mapped to the same dimension space through the feature standardization layer. The time dimension matrix is flattened into a 33x14256 (432x33) vector according to the row, the spatial dimension matrix is flattened into a 33x1089 (33x33) vector, and then the dimensions are unified to 33x1089 through the fully connected network (time side input 14256 dimensions, output 1089 dimensions; space side input 1089 dimensions, output 1089 dimensions). The L2 standardization method is adopted: for the 1089-dimensional feature vector of each node, the L2 norm of the vector is calculated (for example, the L2 norm of the time feature of node 5 = √(0.3²+0.4²+…+0.9²) = 12.5), and each element in the vector is divided by the L2 norm, so that the feature value range after standardization is 0.002-0.072, the time and space feature dimension difference is eliminated, and the fusion of the two in the same scale is ensured. The L2 norm of the standardized feature vector is 1, and the error is ≤0.001.

[0073] Step S32: a cross-modal fusion gate unit is constructed, taking the time-dimension feature vector and the space-dimension feature vector as inputs, and a dynamic fusion weight matrix is generated through a sigmoid activation function, and the element values of the dynamic fusion weight matrix are learned from the correlation strength of the space-time features in different topological states;

[0074] In the embodiment of the application, a cross-modal fusion gate unit is constructed, and a time and space feature vector of 33x1089 is input. The gate unit includes a two-layer fully connected network (input 2178 dimensions, hidden layer 1089 dimensions, and output 1089 dimensions), the time and space feature vector is spliced in the channel dimension to a 33x2178 vector, and then input into the network, and a 33x1089 dynamic fusion weight matrix is generated through a sigmoid activation function (output range 0-1). The element values of the weight matrix are learned from the space-time feature correlation strength: the voltage time sequence trend in the time feature of node 5 and the correlation strength of the node 5-4 correlation weight in the space feature are high, and the corresponding element weight is 0.8; the reactive power fluctuation in the time feature and the correlation strength of the node 5-10 correlation weight in the space feature are low, and the corresponding element weight is 0.3. The gate unit parameters are iteratively optimized through the training data, so that the weight matrix can accurately reflect the space-time correlation in different topological states (such as normal and misconnection of the metering table), in the normal topological state, the average value of the time feature weight is 0.52, and the average value of the space feature weight is 0.48, so as to ensure that the weight distribution conforms to the actual correlation law of the topology.

[0075] Step S33: the time-dimension feature vector and the space-dimension feature vector are nonlinearly coupled based on the dynamic fusion weight matrix, wherein the time-dimension feature vector affects the adjacent aggregation mode of the space-dimension feature vector through the state transmission of the gate recurrent unit, so that the weight of the space adjacency matrix is dynamically adjusted with time, and meanwhile the space-dimension feature vector corrects the time sequence dependence strength of the time-dimension feature vector through the graph attention coefficient, so as to enhance the recognition ability of the space-time coupling anomaly, so as to generate a fused feature vector;

[0076] In the embodiment of the application, the time and space feature vectors are nonlinearly coupled based on a dynamic fusion weight matrix. The time dimension feature vector is transmitted by the state of the gated recurrent unit (forget gate weight 0.6, input gate weight 0.7), and the voltage time sequence change (for example, the voltage rises by 0.2 from t=100 to t=200) is converted into a weight adjustment signal of the spatial adjacency matrix, so that the spatial weight of node 5-4 is dynamically adjusted from 0.73 to 0.81 (matching voltage correlation enhancement); the spatial dimension feature vector corrects the time sequence dependence strength of the time feature by the graph attention coefficient (node 5-4 attention coefficient 0.82), and feeds back the spatial correlation strength of node 5-4 to the time side, so that the voltage time sequence dependence weight from t=200 to t=300 is increased from 0.5 to 0.65, and the recognition ability for voltage abnormal fluctuation (for example, voltage sudden drop caused by account misconnection) is enhanced. The coupled fusion feature vector is 33x1089, and in the fusion feature of node 5, the feature value (0.05-0.072) of the spatiotemporal coupling region is significantly higher than that (0.002-0.03) of the non-coupling region, ensuring that the coupling feature can reflect the spatiotemporal correlation.

[0077] Step S34: The fused feature vector is fused by residual connection and layer normalization processing, the residual connection is used to retain the original spatiotemporal feature information, and the layer normalization is used to stabilize the feature distribution in the training process, to generate a line topology feature representation containing spatiotemporal joint semantic information;

[0078] In the embodiment of the application, the fused feature vector is processed by residual connection and layer normalization. The residual connection adds the original time and space feature vectors to the fused feature vector with a weight of 0.3: fused feature vector = fused feature x 0.7 + time feature x 0.15 + space feature x 0.15, for example, an element in the fused feature of node 5 = 0.06 x 0.7 + 0.05 x 0.15 + 0.04 x 0.15 = 0.057, retaining the original spatiotemporal feature information. The layer normalization calculates the mean (for example, the mean of node 5 is 0.035) and variance (0.0004) of the 1089-dimensional fused feature of each node, and processes it according to the formula (element value-mean) / sqrt(variance+1e-5), so that the normalized feature mean is 0 and the variance is 1, and the feature distribution fluctuation amplitude is reduced from ±0.02 to ±0.005 in the training process, stabilizing the training process. Finally, a 33x1089 line topology feature representation is generated, containing spatiotemporal joint semantic information, for example, the representation of node 5 reflects both the voltage time sequence trend and the change of node correlation weight.

[0079] Step S35: Abnormal feature mining is performed on the line topology feature representation, an autoencoder structure is used to reconstruct the line topology feature representation, a reconstruction error is calculated, an error threshold is set, features with a reconstruction error exceeding the error threshold are determined as abnormal topology features, and a unified topology feature representation containing the abnormal topology is fused and generated.

[0080] In the embodiment of the application, the self-encoder structure (encoder input 1089 dimensions, hidden layer 512 dimensions; decoder input 512 dimensions, output 1089 dimensions) is adopted to reconstruct the line topology feature representation. The 33x1089 feature representation is input into the self-encoder, and the decoder outputs the reconstructed feature vector. The reconstruction error (mean square error, MSE) of each node is calculated: the MSE of the normal topology node (such as node 1) is 0.002; the MSE of the account misconnection node (such as node 8) is 0.015. The error threshold is set to 0.008 (based on the MSE mean value of 1000 normal topology samples + 2 times the standard deviation), and the node features with MSE>0.008 are determined as abnormal topology features (nodes 8, 12, and 5 other nodes). The abnormal and normal features are fused: the abnormal node feature is marked with an abnormal type (account misconnection / short connection), and the normal node feature retains the original information, generating a 33x1089 unified topology feature representation. The abnormal marker bit of the abnormal node in the representation is set to 1, and the normal node is set to 0, ensuring that the abnormal and normal topology features can be intuitively distinguished.

[0081] Further, step S4 includes the following steps:

[0082] Step S41: Constructing a dynamic topology state prediction head, using a graph convolution decoder to reconstruct the topology relationship of the unified topology feature representation. The convolution kernel parameters of the graph convolution decoder are determined by learning the mapping rule of the topology feature and the connection relationship. The output is a connection probability matrix between nodes, and the element value of the connection probability matrix represents the probability of existing effective connection between the corresponding nodes.

[0083] In the embodiment of the application, a dynamic topology state prediction head is constructed, and a 3-layer graph convolution decoder is used to reconstruct the topology relationship of the 33x1089 unified topology feature representation. The convolution kernel parameters of each layer of the graph convolution decoder are determined by learning the mapping rule of the topology feature and the connection relationship: the first layer of the convolution kernel size is 3x3, the input channel is 1089, the output channel is 512, the convolution kernel weight is initialized using the Xavier uniform distribution, and the mapping weight of the voltage time series feature and the node 5-4 connection relationship is learned to be 0.7 through training; the second layer of the convolution kernel size is 3x3, the input channel is 512, the output channel is 256, and the mapping of the power flow feature and the key tie line connection relationship is mainly learned; the third layer of the convolution kernel size is 1x1, the input channel is 256, and the output channel is 33. The output is converted into a probability value of 0-1 through a Softmax activation function, and a 33x33 connection probability matrix is generated. The element value of the matrix represents the effective connection probability between nodes: the element value of node 1-2 is 0.98 (almost certain to be connected), the element value of node 8-9 is 0.25 (high probability of disconnection), and the element value of node 18-19 is 0.92 (normal connection). The matching degree of the connection probability matrix and the actual topology connection relationship is optimized through training iteration, and the final matching accuracy is ≥93%.

[0084] Step S42: design a spatio-temporal joint loss function, which contains a time-dimension time series prediction loss and a space-dimension topological structure loss, the time-dimension time series prediction loss is used to constrain the prediction accuracy of the model on the time series change of the electrical quantity, and the space-dimension topological structure loss is used to optimize the prediction accuracy of the node connection relationship, and a dynamic weighting factor is used to balance the optimization objectives of the two dimensions;

[0085] In the embodiment of the present application, a spatio-temporal joint loss function is designed, which contains a time-dimension time series prediction loss and a space-dimension topological structure loss. The time-dimension time series prediction loss adopts an improved Huber loss to constrain the time series prediction accuracy of the electrical quantity (voltage, current): the deviation between the voltage time series prediction value of node 5 and the true value, small deviation (≤0.12pu) uses square loss (such as deviation 0.08pu, loss 0.0032), large deviation (>0.12pu) uses absolute value loss (such as deviation 0.2pu, loss 0.0168), and the time series prediction loss accounts for 40%. The space-dimension topological structure loss contains a cross-entropy loss (accounting for 40%) and a link prediction loss (accounting for 20%): the cross-entropy loss constrains the node connection state classification (such as node 8-9 prediction error loss 0.0316), and the link prediction loss optimizes the edge weight ranking (such as the actual connection edge prediction score is higher than the non-connection edge by 0.3 or more), and the topological structure loss accounts for 60%. The dynamic weighting factor is adjusted according to the training process: the time series and topological loss weight is 0.5:0.5 in the first 50 training rounds, gradually adjusted to 0.4:0.6 in the 51-150 training rounds, and stabilized at this ratio after the 151 training round, to ensure the balance of the two-dimensional optimization objectives, and the joint loss function value is stabilized at 0.015-0.02 after training convergence.

[0086] Step S43: use an adaptive learning rate optimization algorithm to train the end-to-end model, the initial learning rate is set according to the data size and network complexity, the initial learning rate is periodically adjusted by using a cosine annealing mechanism in the training process, and an early stopping mechanism is introduced, the training is stopped when the validation set loss does not decrease for a plurality of rounds in succession, the model parameters with the optimal performance on the validation set are saved, to construct a dynamic topological state prediction model;

[0087] In the embodiment of the application, the end-to-end model is trained by using an adaptive learning rate optimization algorithm (AdamW). The data scale is 33-node 1-year operation data (17520 hours, 1 sample per hour), the network includes 5 fully connected layers and 3 graph convolution layers, and the initial learning rate is set to 0.005 (calibrated based on the data scale and network complexity to ensure stable initial training loss reduction). In the training process, the learning rate is adjusted by using the cosine annealing mechanism: the period is set to 50 rounds, the learning rate is periodically changed between 0.005 and 0.0001, the learning rate is reduced to 0.0001 at the 50th round, and the learning rate is increased to 0.005 at the 100th round to avoid local optimum. The early stopping mechanism is introduced, and the training is stopped when the validation set loss does not decrease (deviation ≤0.001) for 15 consecutive rounds. When the training is stopped at the 180th round, the validation set loss is stable at 0.017 for 15 consecutive rounds, the training is stopped, and the model parameters (including convolution kernel weights and fully connected layer parameters) of this round are saved. In the model training process, the batch size is 32, the regularization coefficient L2 is 1e-4, and the dropout is 0.3. The final model has a topology connection prediction accuracy of 96.2% on the test set, and the electrical quantity time series prediction error is ≤0.05pu.

[0088] Step S44: The real-time collected line operation data is converted into a unified topology feature representation according to the processing procedures of steps S1-S3, input into the trained dynamic topology state prediction model, and output of the dynamic topology state prediction result, which includes a node connection relationship matrix, a device state health score, and a topology change trend probability distribution.

[0089] In the embodiment of the application, the real-time collected line operation data (voltage, current and power data of 33 nodes within 1 hour, sampling frequency 1 minute / time) is processed according to steps S1-S3: step S1 collects data and generates initial topology data, step S2 pre-processes (interpolates and fills in 2 missing values, and corrects 1 abnormal account), and step S3 generates a unified topology feature representation of 33x1089 through a space-time separation encoder and feature fusion. The feature representation is input into the trained dynamic topology state prediction model, and the model outputs the dynamic topology state prediction result: a 33x33 node connection relationship matrix (node 1-2 connection probability 0.99, node 8-9 0.26, node 18-19 0.91); a state health score of 33 devices (node 8 device 72 points, node 18 device 88 points, node 1 device 95 points); and a topology change trend probability distribution (future 1 hour node 8-9 disconnection probability 78%, node 18-19 tripping probability 12%, and overall topology stability probability 85%). The prediction result output frequency is consistent with the data collection frequency, which is updated once per hour to ensure real-time reflection of the topology state.

[0090] Step S45: post-processing the dynamic topology state prediction result, verifying the dynamic topology state prediction result in combination with historical topology data, ensuring the reliability of the output result, and thus outputting the real-time topology connection relationship and device state change trend corresponding to the power distribution network; and performing dynamic topology reconstruction based on the real-time topology connection relationship and device state change trend corresponding to the power distribution network, to generate a current-time reliable topology graph containing potential branch disconnection and temporary trip.

[0091] In the embodiment of the application, the dynamic topology state prediction result is post-processed, the connection relationship matrix is converted into a binary topology graph (connected 1, disconnected 0) by threshold determination (connection probability ≥ 80% is determined to be connected, ≤ 30% is determined to be disconnected), and a device health score ≤ 75 is marked for abnormal early warning. In combination with historical topology data (topology connection records and device health change in the same time period in the past one month), it is verified that node 8-9 has been disconnected for 3 times in the same time period in the past one month, the current prediction disconnection probability 78% matches the historical regularity; node 18-19 has no trip record in the past one month, and the current prediction trip probability 12% conforms to the historical stability, the output result is verified to be reliable, the real-time topology connection relationship (32 connected branches, 1 disconnected branch) and the device state change trend (the device health of node 8 is continuously decreasing, and the device of node 18 is stable) of the power distribution network are generated. Based on the result, dynamic topology reconstruction is performed, potential branch disconnection (node 8-9) and normal branch are marked, a current-time reliable topology graph (containing 32 normal branches and 1 marked disconnected branch) is generated, and a device state early warning list is attached, which is used for operation and maintenance scheduling.

[0092] Further, the time dimension time sequence prediction loss in step S42 adopts an improved Huber loss function, and the topology structure loss in the spatial dimension adopts a combination loss mechanism, and the specific implementation process includes:

[0093] In the time dimension loss calculation, the improved Huber loss function realizes the differentiated processing of different amplitude electrical quantity fluctuations through a dynamic threshold adjustment module, adopts a square loss for small amplitude normal fluctuations to improve the prediction accuracy, and adopts an absolute value loss for large amplitude fault fluctuations to enhance the anti-interference ability of the model, and the dynamic threshold is adaptively adjusted according to historical electrical quantity fluctuation data and fault statistical regularity;

[0094] In the embodiment of the application, the deviation of the time series prediction value of the electrical quantity of the 33 nodes from the true value (such as voltage and power) is processed by a dynamic threshold adjustment module in the time dimension loss calculation by using an improved Huber loss function. The dynamic threshold is adaptively adjusted based on the historical electrical quantity fluctuation data (normal fluctuation range 0-0.12pu, fault fluctuation >0.12pu) and the fault statistical law (fault fluctuation accounts for 5%) in the past one year, and is set to 0.12pu. For small amplitude normal fluctuation (deviation ≤0.12pu), square loss calculation is used: the voltage prediction deviation of node 5 is 0.08pu, and the loss value = 0.5×(0.08)²=0.0032, which improves the normal time series prediction accuracy; for large amplitude fault fluctuation (deviation >0.12pu), absolute value loss calculation is used: the voltage prediction deviation of node 8 caused by the misconnection of the account is 0.2pu, and the loss value = 0.12×0.2-0.5×(0.12)²=0.024-0.0072=0.0168, which enhances the anti-interference ability of the model to the fault fluctuation. The dynamic threshold is updated daily to ensure that it matches the real-time electrical quantity fluctuation characteristics, and the threshold is adjusted to 0.15pu during the peak load period to avoid misjudgment of normal load fluctuation as a fault, and the time dimension loss mean is controlled in the range of 0.005-0.018, which takes into account the loss calculation accuracy of normal and fault scenarios.

[0095] Preferably, in the spatial dimension loss calculation, cross-entropy loss is used to constrain the classification accuracy of the node connection state, and a weighting factor is used to highlight the prediction accuracy of the connection relationship of the key tie lines and the substation outgoing lines corresponding nodes, and to reduce the influence of the prediction error of the non-key lines on the overall topology analysis result;

[0096] In the embodiment of the application, in the spatial dimension loss calculation, cross-entropy loss is used to constrain the prediction accuracy of the node connection state (1 for connected, 0 for not connected) of the 33×33 spatial adjacency matrix. A weighting factor is used to highlight the key lines: the weighting factor of the key tie lines (such as nodes 1-2 and nodes 5-4, which account for more than 50% of the power transmission in the region) is set to 0.8; the weighting factor of the substation outgoing lines (nodes 1-3 and nodes 1-5) is set to 0.9; the weighting factor of the non-key lines (such as nodes 25-26, which account for less than 5% of the power) is set to 0.3. The cross-entropy loss is calculated: the connection state prediction of node 1-2 is correct (true 1, predicted 1), and the loss value =-0.9×1×ln(0.95)≈0.046 (prediction probability 0.95); the connection state prediction of node 8-9 is wrong (true 0, predicted 1) due to the misconnection of the account, and the loss value =-0.3×0×ln(0.1)-0.3×1×ln(0.9)≈0-(-0.0316)=0.0316, which reduces the influence of the prediction error of the non-key lines on the overall topology. After balancing the total loss in the spatial dimension by using the weighting factor, the loss proportion of the key lines is 65%, and the loss proportion of the non-key lines is 35%, which ensures the prediction accuracy of the core topology connection.

[0097] Preferably, link prediction loss is introduced, and a Bayesian personalized ranking loss function is used to compare the predicted scores of actual connection edges and non-connection edges, so as to optimize the relative ranking relationship of edge weights, and enable the model to more accurately identify effective connections and ineffective connections in the topology structure.

[0098] In the embodiment of the present application, by introducing link prediction loss, a Bayesian personalized ranking loss function is used to compare the predicted scores of actual connection edges (32) and non-connection edges (33*33-32=1057) in the spatial adjacency matrix. For each actual connection edge, 5 non-connection edges are randomly selected to form a comparison group, and the loss is calculated: the predicted score of the actual connection edge node 5-4 is 0.81, the predicted score of the non-connection edge node 5-9 is 0.2, and the loss value = ln(1+exp(0.2-0.81))=ln(1+0.543)=0.434; the predicted score of the actual connection edge node 1-2 is 0.92, and the predicted score of the non-connection edge node 1-20 is 0.15, and the loss value = ln(1+exp(0.15-0.92))=ln(1+0.463)=0.381. By optimizing the relative ranking of the edge weights through the loss, the predicted score of the actual connection edge is higher than that of the non-connection edge (the difference is ≥0.3), and the model can accurately identify effective connections (such as node 5-4) and ineffective connections (such as node 8-9). The average link prediction loss is controlled in 0.35-0.45, and the edge weight ranking accuracy is ≥92%.

[0099] Preferably, a loss balancing mechanism is constructed to dynamically adjust the weighting factor to weight and fuse the time loss and the space loss, balance the weights of the two at the initial stage of training, and gradually increase the proportion of the space topology loss with the training process.

[0100] In the embodiment of the present application, by constructing a loss balancing mechanism, a weighting factor is adjusted by a dynamic weight to fuse the time loss (Lt) and the space loss (Ls, including cross-entropy loss and link prediction loss). In the early stage of training (1-50 rounds), the weighting factor is set as Lt:Ls=0.5:0.5, which ensures the synchronous learning of space-time features, such as Lt=0.008 and Ls=0.04 in the 10th round, and the total loss = 0.5*0.008+0.5*0.04=0.024; with the training process (51-150 rounds), the proportion of space topology loss is gradually increased, and the weight of Ls is increased by 0.05 and the weight of Lt is reduced by 0.05 every 10 rounds; in the later stage of training (151-200 rounds), the weighting factor is stable as Lt:Ls=0.2:0.8, which preferentially ensures the accuracy of topology structure prediction, such as Lt=0.005 and Ls=0.02 in the 180th round, and the total loss = 0.2*0.005+0.8*0.02=0.017. The dynamic weight is adjusted by the feedback of the topology prediction accuracy rate of the validation set, and when the accuracy rate is greater than or equal to 95%, the weight of Ls is no longer increased, so as to ensure that the model focuses on topology accuracy optimization in the later stage of training, and the final topology prediction accuracy rate is stable at more than 96%.

[0101] Preferably, the loss function is regularized, an L2 regularization term is introduced to suppress the model parameters from being too large, and a Dropout technique is used to randomly shield part of the neurons to prevent model overfitting, so as to design a corresponding space-time joint loss function.

[0102] In the embodiment of the present application, the loss function is regularized, an L2 regularization term and a Dropout technique are introduced to prevent overfitting. The coefficient of the L2 regularization term is set to 1e-4, and the sum of squares of all parameters of the calculation model (such as the weights of the fully connected network and the attention coefficient) is calculated: assuming that the total number of model parameters is 1e5 and the average of the sum of squares of the parameters is 0.02, the L2 regularization loss = 1e-4*0.02 = 0.002, which suppresses the parameters from being too large (the absolute value of a single parameter is less than or equal to 0.5). The Dropout technique is used to randomly shield 30% of the neurons in the fully connected layer and the attention layer, such as 1089 neurons in the time feature processing layer of node 5, of which 327 are randomly shielded, so that the model does not depend on specific neurons and the generalization ability is enhanced. The finally designed space-time joint loss function = dynamic weighted fusion loss + L2 regularization loss, such as dynamic weighted fusion loss 0.015 and L2 loss 0.002 in the 200th round, and the total loss = 0.017. Through the regularization processing, the topology prediction accuracy rate of the model on the test set is improved by 4% compared with that without regularization, the overfitting degree (the difference between the accuracy rates of the training set and the test set) is reduced from 12% to 5%, and the stable generalization of the model is ensured.

[0103] Further, the post-processing of the dynamic topology state prediction result in step S45 includes:

[0104] The basic connectivity indexes of the topology structure are calculated based on the prediction result of the dynamic topology state, including node degree distribution, average path length and clustering coefficient.

[0105] In the embodiment of the application, the basic connectivity indexes of the topology structure are calculated based on the prediction result of the dynamic topology state (a 33*33 space adjacency matrix containing 32 effective connection edges). The node degree distribution is calculated by a degree statistical tool: the root node 1 is connected with 5 lines (degree 5), the key tie node 5 is connected with 3 lines (degree 3), and the terminal node 33 is connected with 1 line (degree 1), thereby generating a degree distribution [5, 3, 2, …, 1], and the average degree is 1.94 (64 / 33). The average path length is calculated by a shortest path algorithm (Dijkstra algorithm) to calculate the shortest path of all node pairs: the node 1 to the node 33 needs to pass through 8 nodes (path length 8), and the node 5 to the node 20 needs to pass through 5 nodes (path length 5), the path lengths of all 33*32 / 2=528 groups of node pairs are counted, and the average path length = total path length / 528 = 1260 / 528≈2.39. The clustering coefficient is calculated according to the formula: the actual connection edge number between the 3 neighbor nodes (4, 6, 10) of the node 5 is 2 (4-6, 6-10), the theoretical maximum connection edge number is 3 (3*2 / 2), and the clustering coefficient = 2 / 3≈0.67; the average clustering coefficient of all nodes = total clustering coefficient / 33 = 15.2 / 33≈0.46, and the basic connectivity indexes completely reflect the connection density and path characteristics of the topology.

[0106] Preferably, a multi-scenario fault set is generated by Monte Carlo simulation to simulate the influence of different types of faults on the topology structure, and the load transfer margin and power supply recovery time efficiency under different fault scenarios are calculated to generate the dynamic response capability indexes of the topology structure.

[0107] In the embodiment of the application, 1000 groups of multi-scenario fault sets are generated by Monte Carlo simulation, covering line circuit breaking (accounting for 60%), node fault (accounting for 30%), and device malfunction (accounting for 10%). Simulate line 1 (node 1-2) circuit breaking fault: calculate load transfer margin, transfer 20MW load of line 1 to adjacent line 3 (node 1-3), the rated capacity of line 3 is 25MW, the current load is 15MW, and the margin = 25-15-20 =-10MW (need to be further transferred to line 5); the power supply recovery time is accumulated by fault positioning (5 seconds), isolation (10 seconds), and transfer (20 seconds), and the recovery time is 5+10+20 = 35 seconds. Simulate node 10 fault: load transfer margin = sum of remaining capacities of surrounding lines 6 (node 5-6) and 9 (node 9-10) 30MW-node 10 load 15MW = 15MW; recovery time = 8+12+18 = 38 seconds. Statistically, the average load transfer margin of 1000 groups of fault scenarios is 8.5MW, the average power supply recovery time is 32 seconds, the dynamic response capability index is generated, and the larger the index value is, the stronger the response capability of the topology to the fault is.

[0108] Preferably, a fault propagation model is constructed to simulate the influence range of a single node fault on the global topology based on the node electrical coupling relationship and the topology connection relationship, and to calculate a fault propagation index to generate an anti-disturbance capability index of the topology structure;

[0109] In the embodiment of the application, a fault propagation model is constructed to simulate fault propagation based on the node electrical coupling relationship (voltage coupling coefficient ≥ 0.8) and the topology connection relationship (directly adjacent). Take node 8 fault (accounting error leading to voltage drop) as an example: node 8 is directly adjacent to nodes 6, 7, and 9 (topology connection), and is electrically coupled to nodes 5 and 10 (voltage coupling coefficient 0.82, 0.78), the fault first propagates to the directly adjacent nodes (voltage drop of 6, 7, and 9 is 0.15pu), and then spreads to the electrically coupled nodes (voltage drop of 5 and 10 is 0.1pu). Calculate the fault propagation index = number of propagation nodes / total number of nodes x 100%+maximum voltage drop / initial voltage x 100%, the number of propagation nodes is 6 (8, 6, 7, 9, 5, 10), the total number of nodes is 33, the maximum voltage drop is 0.2pu, and the initial voltage is 1.0pu, the propagation index = 6 / 33 x 100%+0.2 / 1.0 x 100%≈18.18%+20%=38.18%. Simulate node 1 fault (root node): the number of propagation nodes is 33, the maximum voltage drop is 0.5pu, and the propagation index = 100%+50%=150%. The average of all node fault propagation indexes is 45.2%, an anti-disturbance capability index is generated, and the lower the index is, the stronger the anti-fault propagation capability of the topology is.

[0110] Preferably, a topology reliability scoring system is established, and the basic connectivity index, dynamic response capability index and disturbance resistance capability index are fused according to weights to generate a comprehensive reliability score of the topology structure.

[0111] In the embodiment of the application, the topology reliability scoring system is established, and the basic connectivity index (weight 0.3), the dynamic response capability index (weight 0.4) and the disturbance resistance capability index (weight 0.3) are fused according to linear weighting. The basic connectivity index is standardized: the degree distribution mean value is 1.94 (100 points correspond to 2.0, and 97 points are scored), the average path length is 2.39 (100 points correspond to 2.0, and 83 points are scored), and the clustering coefficient is 0.46 (100 points correspond to 0.5, and 92 points are scored), and the mean value is (97+83+92) / 3=90.67, and the weighted score is 90.67*0.3≈27.2. The dynamic response capability index is standardized: the load transfer margin is 8.5 MW (100 points correspond to 10 MW, and 85 points are scored), and the recovery time is 32 seconds (100 points correspond to 30 seconds, and 93 points are scored), and the mean value is (85+93) / 2=89, and the weighted score is 89*0.4≈35.6. The disturbance resistance capability index is standardized: the propagation index is 45.2% (100 points correspond to 30%, and 66 points are scored), and the weighted score is 66*0.3≈19.8. The comprehensive reliability score is 27.2+35.6+19.8=82.6 (100 points), and the higher the score, the stronger the topology reliability.

[0112] Preferably, potential fault links with a comprehensive reliability score lower than a set threshold are screened out, root cause analysis is performed in combination with historical topology data, electrical characteristics and topology characteristics of the potential fault links are extracted, similarity matching is performed with historical fault cases in the historical topology data, the fault cause is determined, and alarm information containing a fault position, possible causes and an impact range and topology repair suggestions are output; the dynamic topology state prediction result is verified and predicted based on the corresponding fault cause, and real-time topology connection relationships and device state change trends of the power distribution network are output.

[0113] In the embodiment of the application, by setting the comprehensive reliability score threshold of 75 points, the line 8 (node 8-9) with a score of 72 points and the line 15 (node 15-16) with a score of 68 points are screened out as potential fault links. Combined with the historical topology data (100 fault cases in the past 3 years), the root cause analysis is carried out: the electrical characteristics (voltage fluctuation 0.2pu, power flow deviation 15%) and the topology characteristics (accounting error, adjacent node degree 3) of the line 8 are extracted, which match the historical case "line X accounting error" (similarity 92%), and it is determined that the fault reason is accounting error; the characteristics (impedance increase 20%, connection node degree 2) of the line 15 match the historical case "line Y joint oxidation" (similarity 88%), and it is determined that the fault reason is joint oxidation. The output alarm information is: "fault position: line 8 (node 8-9), possible reason: accounting error, influence range: 6 node voltage drop; fault position: line 15 (node 15-16), possible reason: joint oxidation, influence range: 3 node power anomaly", and the repair suggestion is: "correct the line 8 account to node 8-8, replace the line 15 joint". Based on the fault reason verification prediction: simulate the line 8 score after the accounting correction to improve to 85 points, verify the prediction result is accurate, finally output the real-time topology connection relationship (32 effective edges) and the device state change trend (line 8, 15 need to be monitored) of the distribution network.

[0114] Further, the dynamic topology reconstruction based on the real-time topology connection relationship and the device state change trend of the distribution network in step S45 comprises:

[0115] extracting the real-time topology connection relationship and the device state change trend, wherein the real-time topology connection relationship comprises connection probability values between nodes, and the device state change trend comprises a device health score change curve over time and an abnormal state early warning mark;

[0116] In the embodiments of the present application, the real-time topological connection relationship and the device state change trend of the IEEE33 node power distribution network are extracted. The real-time topological connection relationship calculates the connection probability value between nodes through the dynamic topological state prediction result: for each branch, based on the topological prediction result (connection state 1 is connected, 0 is not connected) in the last 1 hour, 12 time steps, the connection probability value = connected times / total time steps, line 8 (node 8-9) is connected in only 3 of the last 12 steps, the connection probability value = 3 / 12 = 25%; line 1 (node 1-2) is connected in all 12 steps, the connection probability value = 100%; line 15 (node 15-16) is connected in 8 steps, the connection probability value = 8 / 12 ≈ 66.7%, and a 33x33 connection probability matrix is generated. The device state change trend is obtained through the device health score system: the health score is calculated based on the device operating parameters (temperature, humidity, insulation resistance), and the full score is 100 points. The line 8 device health score curve in the last 12 hours is “90→85→80→72” (recorded once every 2 hours), and an abnormal state warning mark (red warning) is marked when the score drops to 75 points; the line 15 health score curve is “92→90→88→85”, and there is no warning mark; the line 1 health score is stable at 95 points, and there is no warning, ensuring that the device state change trend fully reflects the device degradation process and abnormal warning.

[0117] Preferably, a dynamic topological reconstruction rule library is constructed, and for the potential branch disconnection scenario, a double determination condition of connection probability value lower than disconnection determination threshold and adjacent node electrical quantity mutation is set; for the temporary tripping scenario, a joint determination condition of device health score sudden drop, corresponding branch current sudden drop to zero and topological connection relationship matrix of the branch connection probability instantaneous zero is set;

[0118] In the embodiment of the application, two types of scene judgment conditions are set by constructing a dynamic topology reconstruction rule base. The potential branch disconnection scene: the disconnection judgment threshold is set to 30% (based on historical disconnection case statistics, when the connection probability is less than or equal to 30%, the branch disconnection probability is greater than or equal to 90%), and the double judgment condition is “connection probability value ≤ 30% and adjacent node electrical quantity mutation (voltage deviation ≥ 0.1pu or current deviation ≥ 20%)”, for example, the connection probability of line 8 is 25% ≤ 30%, the voltage deviation of adjacent nodes 8 and 9 is 0.15pu ≥ 0.1pu, which meets the condition. Temporary tripping scene: the joint judgment condition is “device health score drops sharply (decreases by ≥ 10 points within 1 hour), corresponding branch current suddenly changes to zero (current value ≤ 0.5A), and the topology connection relationship matrix of the branch connection probability is temporarily zero (decreases from ≥ 80% to ≤ 10% within 1 time step)”, for example, the simulation line 5 (nodes 5-6) trips, the health score decreases from 90 points to 78 points within 1 hour (decreases by 12 points), the branch current decreases from 180A to 0.2A, and the connection probability decreases from 95% to 5%, which meets the condition. All thresholds in the rule base are calibrated by 500 groups of fault cases in the past 3 years to ensure that the accuracy of the judgment condition is ≥ 95%, avoiding misjudgment and omission.

[0119] Preferably, the real-time topology connection relationship is analyzed for topology structure, the connection relationship between nodes is traversed based on the connectivity analysis algorithm in graph theory, the branch with a connection probability value meeting the double judgment condition is marked, and the branch meeting the joint judgment condition is distinguished in combination with the change curve in the device state change trend and the abnormal state early warning mark, so as to preliminarily locate the fault position of the potential branch disconnection and temporary tripping;

[0120] In the embodiment of the application, by analyzing the 33x33 connection probability matrix of the real-time topological connection relationship, a depth-first search algorithm in graph theory is used to traverse all the connection relationships between nodes (starting from root node 1, adjacent nodes 2, 3, 5, etc. are traversed in turn, and the connection probability value of each branch is recorded). Mark the branch that meets the double judgment condition of potential branch disconnection: traverse line 8 (connection probability 25%≤30%, voltage deviation 0.15pu≥0.1pu), line 22 (nodes 22-23, connection probability 28%≤30%, current deviation 22%≥20%), combined with the device state change trend, the line 8 health score curve shows a downward trend and abnormal early warning, the line 22 health score is stable but there is an electrical quantity mutation, and the two branches are preliminarily located as potential branch disconnection fault positions. Determine the branch that meets the temporary trip joint determination condition: traverse line 18 (nodes 18-19) connection probability from 92% to 8%, device health score from 88 points to 75 points (decrease of 13 points) within 1 hour, branch current from 150A to 0.3A, meet the joint determination condition, and preliminarily locate as a temporary trip fault position. All positioning results are compared with the fault alarm information of the real-time monitoring system, and the verification accuracy is ≥92%, ensuring accurate fault location.

[0121] Preferably, a multi-source data cross-validation mechanism is introduced to compare the preliminarily located fault position with the switch state signal and fault recording data collected by the real-time monitoring system of the distribution network. If the connection state of the fault position is consistent with the actual on-off state of the switch, and the electrical quantity change matches the fault recording data, the abnormal scenario is confirmed. If there is data deviation, a backtracking verification process is started to retrieve the unified topological feature representation generated in step S3 again, and the connection probability value between each node is corrected.

[0122] In the embodiment of the application, the preliminary fault location is compared with the data of the real-time monitoring system of the power distribution network by introducing a multi-source data cross-validation mechanism. For line 8 (potential disconnection): the switch state signal of the monitoring system is called, and the switch corresponding to line 8 is actually in the open state (consistent with the disconnection trend reflected by the connection probability of 25%); the fault recording data is checked, and the voltage of adjacent nodes 8 and 9 of line 8 suddenly drops from 0.98pu to 0.83pu at 10:00 (deviation 0.15pu, matching the sudden change of electrical quantity), confirming the potential branch disconnection scenario. For line 22 (potential disconnection): the switch state shows that the switch is closed, but the current of line 22 in the fault recording data decreases from 120A to 95A (deviation 20.8%, meeting the mutation condition), there is data deviation, and the backtracking verification process is started. The unified topology feature representation (spatiotemporal fusion feature vector of line 22) generated in step S3 is called again, and the connection probability calculation is corrected: the prediction data of the last 30 minutes (4 steps connected, 2 steps not connected) is added, the new connection probability value=(28%×12+4 / 6×6) / (12+6)=(3.36+4) / 18≈40.9% (higher than the threshold of 30%), and the potential disconnection is excluded. For line 18 (temporary trip): the switch state is open, and the fault recording shows that the current suddenly changes to 0.3A, which completely matches the determination condition, confirming the temporary trip scenario.

[0123] Preferably, a topology completion algorithm is used to repair the topology structure of the confirmed abnormal scenario. For potential branch disconnection, a virtual backup connection path is generated according to the historical connection mode of the disconnected branch and the topology association characteristics of the surrounding nodes; for temporary trip, a temporary replacement branch is marked based on the historical recovery record of the tripped device and the redundant design of the power grid topology, thereby generating a backup path.

[0124] In the embodiment of the present application, the confirmed abnormal scenarios are repaired by using a topology completion algorithm. For line 8 (potential disconnection): extract its historical connection mode (main connection node 8 and node 7 for nearly 6 months of normal operation, backup connection node 8 and node 10), combine the topological association characteristics of the surrounding nodes (node 7 connects nodes 5 and 6, and node 10 connects nodes 5 and 9), supplement the virtual backup connection path "node 8→node 7→node 5→node 9" (instead of the original "node 8→node 9" path), and the total impedance of the path is 0.35Ω (close to the original path 0.32Ω). For line 18 (temporary trip): query the historical recovery record of the tripping device (the node 17→node 19 path is temporarily recovered after nearly 3 trips), combine the redundant design of the power grid topology (there is a backup line between node 17 and node 19, and the current load rate is 60%), and mark "node 18→node 17→node 19" as a temporary replacement branch. The rated current of this branch is 200A, the current is 120A, and it meets the load transfer demand (the original load of line 18 is 80A). After repair, all backup paths are verified by power flow calculation to ensure that the power transmission loss is ≤5% and the voltage drop is ≤0.05pu.

[0125] Preferably, a visualization generation module of the current time topology graph is constructed, based on the device spatial coordinates of the power distribution network geographic information system, different colors are used to identify normal branches, potential disconnection branches, and tripped branches, and the position information, abnormal determination basis, and recommended treatment scheme of the abnormal branch are marked in the current time topology graph, including repair priority and backup path;

[0126] In the embodiment of the present application, a visualization generation module of the current time topology graph is constructed, based on the device spatial coordinates of the power distribution network geographic information system (for example, the coordinates of node 8 are 30.1256°N, 120.5701°E, and the coordinates of node 18 are 30.1302°N, 120.5789°E), different colors are used to identify the state of the branch: normal branches (such as lines 1 and 15) are green, potential disconnection branches (line 8) are yellow, and tripped branches (line 18) are red. In the topology graph, the information of the abnormal branch is marked: line 8 is marked with "position: node 8-9, determination basis: connection probability 25%+voltage deviation 0.15pu, repair priority: high, backup path: 8→7→5→9"; line 18 is marked with "position: node 18-19, determination basis: health degree drop 13 points+current zero, repair priority: medium, backup path: 18→17→19". The visualization interface supports zooming (scale 1:10000) and clicking to view details (clicking on the line to display real-time current and voltage), ensuring that the operation and maintenance personnel can intuitively obtain abnormal information and treatment schemes.

[0127] Preferably, the current time topology is verified for reliability, the connectivity, redundancy and load coverage capability indexes of the current time topology are calculated, if the connectivity is lower than the power grid safe operation threshold, the redundancy cannot meet the fault transfer requirement or the load coverage capability index has a blind area, then the standby path is returned to be optimized again, if all the reliability indexes meet the requirements, then the final current time reliable topology is output.

[0128] In the embodiment of the application, the core indexes are calculated by verifying the reliability of the current time topology. The connectivity is calculated by graph theory algorithm: all the 33 nodes of the power grid are connected to the root node 1 through at least one path, the connectivity = 100% (the power grid safe operation threshold ≥ 95%). The redundancy is calculated by the number of standby paths of each branch: the normal branch has an average of 2 standby paths, the abnormal repair branch (lines 8, 18) has one standby path, the redundancy average is 1.8 (the fault transfer requirement threshold ≥ 1.2). The load coverage capability index is calculated by the load point connectivity statistics: all the 120 load points (33 nodes each containing 3-4 load points) of the power grid can obtain power supply through the main path or the standby path, and there is no coverage blind area (the load coverage capability threshold ≥ 100%). All the indexes meet the requirements, the standby path does not need to be optimized, and the final current time reliable topology is output. The topology includes 32 normal branches and 2 repair branches, and can be directly used for power distribution network operation and maintenance scheduling.

[0129] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the application.

[0130] The above description is only a specific embodiment of the application, which enables those skilled in the art to understand or implement the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A space-time fusion neural network line topology analysis method for power distribution networks, characterized in that, The method comprises the following steps: Step S1: obtaining original line topology data corresponding to the power distribution network, including account information corresponding to line operation, electrical quantity time series data and equipment spatial distribution data, and structurally disassembling and preprocessing the original line topology data through a space-time separation encoder to construct a space-time double graph structure including a time series tensor and a spatial adjacency matrix, wherein the time series tensor records dynamic evolution characteristics of voltage and current corresponding electrical quantities, and the spatial adjacency matrix represents physical connection relationships and initial topological correlation strength of equipment; Step S2: constructing a bidirectional dynamic feature interaction mechanism based on the space-time double graph structure, to model long-term dependence of the time series tensor through a bidirectional gate graph recurrent unit in the time dimension, capture bidirectional evolution rules of electrical quantities at multiple time points, and extract multi-scale topological features of the spatial adjacency matrix through a hierarchical graph attention pooling module in the spatial dimension, to establish dynamic correlation weights between nodes, and generate a set of space-time separated feature vectors, including time dimension feature vectors and spatial dimension feature vectors; Step S3: using a cross-modal gate cycle and graph attention hybrid mechanism to deeply couple and fuse the set of space-time separated feature vectors, and realizing dynamic weighted interaction of the time dimension feature vectors and the spatial dimension feature vectors through learnable gate weight parameters, to generate a unified topological feature representation including an abnormal topological representation; Step S4: constructing a dynamic topological state prediction model based on the unified topological feature representation, to optimize a space-time joint loss function through end-to-end training, and output real-time topological connection relationships and equipment state change trends corresponding to the power distribution network, and simultaneously reconstruct a dynamic topology based on the real-time topological connection relationships and the equipment state change trends corresponding to the power distribution network, to generate a reliable topological graph at the current time point including potential branch disconnection and temporary tripping.

2. The power distribution network-oriented spatio-temporal fusion neural network line topology analysis method according to claim 1, characterized in that, Step S1 comprises the following steps: Step S11: cooperatively collecting original line topology data corresponding to the power distribution network through a power distribution network monitoring system, an equipment account management platform and a geographic information system, wherein the account information corresponding to line operation includes line type, connection node number and laying method information, the electrical quantity time series data includes minute-level sampled three-phase voltage, current effective value and active and reactive power data, and the equipment spatial distribution data includes latitude, longitude and elevation data of transformer substations, switch stations and line towers; Step S12: preprocessing the original line topology data, to fill in missing values in the electrical quantity time series data and the equipment spatial distribution data using an interpolation algorithm, and identifying and correcting abnormal records corresponding to line misconnection and missing connection points in the account information using an isolation forest algorithm, and simultaneously eliminating the influence of electrical parameters with different dimensions through a range normalization method, to generate line topology standardized data; Step S13: input the line topology standardized data into a space-time separation encoder, and structurally disassemble the line topology standardized data in the time dimension and the space dimension to generate a time series tensor indexed by a time stamp and a space adjacency matrix with a node-edge relationship as a core, wherein a dimension of the time series tensor is determined by a sampling time length and a number of electrical parameter types, and an initial value of an element of the space adjacency matrix is calculated from a physical distance between devices and a line impedance; Step S14: dynamically correcting the space adjacency matrix to adjust a weight of the initial value of the element in the space adjacency matrix by introducing real-time operation parameters of the line, including a power flow distribution and a voltage drop rate, so that the space adjacency matrix reflects the physical connection relationship and the electrical function coupling strength of the devices, and a space-time double graph structure is constructed and generated in combination with the time series tensor.

3. The power distribution network-oriented spatio-temporal fusion neural network line topology analysis method according to claim 2, characterized in that, The space-time separation encoder in step S13 includes a feature mapping layer, a dimension alignment module, and a space-time consistency verification unit, and the specific implementation process includes: Step S131: the feature mapping layer structurally disassembles the line topology standardized data in the time dimension and the space dimension by a multi-modal embedding function to generate a time series tensor indexed by a time stamp and a space adjacency matrix with a node-edge relationship as a core; Step S132: the dimension alignment module synchronizes time axes of different sampling frequency data by a dynamic time warping algorithm, processes high-frequency sampled time series tensors by a sliding window mean frequency reduction, and processes low-frequency updated space adjacency matrices by linear interpolation frequency increase, so that the time series tensor and the space adjacency matrix are consistent in the time scale, and feature misalignment caused by differences in data sampling frequency is avoided; Step S133: the space-time consistency verification unit identifies asynchronous deviations of time series data and space data by calculating a matching degree of a time stamp of a mutation time corresponding to the time series tensor and a time stamp of each topology structure change record in the space adjacency matrix, and starts a secondary verification mechanism for samples with asynchronous deviations exceeding a set threshold, to correct the asynchronous deviations by combining historical topology evolution rules and electrical quantity change correlation rules, to generate processed time series tensors and space adjacency matrices; Step S134: dimension compression and feature screening are performed on the processed time series tensors and the space adjacency matrices to extract key time series features in the time series tensors by principal component analysis technology, and to simplify the space adjacency matrices by a sparse matrix decomposition method, so as to reduce data calculation complexity while retaining core information.

4. The power distribution network oriented spatio-temporal fusion neural network line topology analysis method according to claim 1, characterized in that, The bidirectional gated graph recurrent unit in step S2 includes a forward propagation chain, a backward propagation chain, a state fusion gate, and a residual connection module, and the bidirectional gated graph recurrent unit through the time dimension models long-term dependence of the time series tensor, including: The forward propagation chain encodes the time series tensor from a historical time to a current time to generate a current time forward state vector by filtering current time new features through an input gate, discarding redundant historical information through a forget gate, and generating the current time forward state vector through an output gate, wherein gated weight parameters are determined by electrical quantity features at the current time and historical state features. The backward propagation chain decodes the time series tensor from the future time to the current time to generate a current time reverse state vector by using a gating structure symmetrical to the forward propagation chain, wherein the gating weight parameters are determined by learning the future time electrical quantity change trend; The state fusion gate dynamically weights and fuses the forward state vector and the reverse state vector through an attention mechanism, calculates the similarity of the forward state vector and the reverse state vector in different feature dimensions, assigns a fusion weight according to the similarity, and generates a comprehensive state vector containing bidirectional time series information; The residual connection module element-wise adds the original time series tensor and the comprehensive state vector, and relieves the gradient vanishing problem through layer normalization processing. Meanwhile, the attention mechanism is introduced to give different weights to time series features of different electrical parameter types, so as to capture the bidirectional evolution law of electrical quantities at multiple times, and generate a time dimension feature vector.

5. The power distribution network oriented spatio-temporal fusion neural network line topology analysis method according to claim 1, characterized in that, The hierarchical graph attention pooling module in step S2 includes multiple levels of attention heads, feature enhancement units and topology robustness correctors. The multi-scale topology feature extraction of the spatial adjacency matrix by the hierarchical graph attention pooling module in the spatial dimension includes: The bottom layer attention head focuses on the local topology features of the directly adjacent nodes, calculates the electrical quantity similarity between nodes and the physical distance attenuation coefficient, and weights and fuses the two as the bottom layer attention weight; The middle layer attention head captures the intermediate scale topology features of the 2-N order neighborhood through jump connection. The receptive field range of each level of attention head expands exponentially, and the cross-layer residual connection is introduced to prevent feature degradation. The middle layer attention weight is calculated by the local topology features and the global topology structure features. The high layer attention head extracts the global topology features of the power distribution network through global pooling operation, and gives different weights to the topology features of different regions according to the tree structure characteristics of the power grid topology and the power flow distribution law. The feature enhancement unit performs channel attention weighting on the feature channels output by each level of attention head, calculates the contribution degree of different feature channels to topology state recognition, and weights and corrects the topology feature information of each level of attention head based on the contribution degree of different feature channels to topology state recognition. The topology robustness corrector adjusts the weight proportion of the abnormal connection corresponding to the abnormal connection containing the account error connection and the short connection by calculating the feature consistency score. When it is detected that the topology features of the line are seriously inconsistent with the surrounding lines, the weight of the line in the spatial adjacency matrix is automatically reduced to establish the dynamic association weight between nodes and generate a spatial dimension feature vector.

6. The power distribution network oriented spatio-temporal fusion neural network line topology analysis method according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Extract the generated time dimension feature vector and spatial dimension feature vector, map the two types of features to the same dimensional space through the feature standardization layer, and eliminate the feature dimension difference by using the L2 standardization method to ensure that the space-time features are fused in the same scale. Step S32: A cross-modal fusion gating unit is constructed, taking the time-dimension feature vector and the space-dimension feature vector as inputs, and generating a dynamic fusion weight matrix through a sigmoid activation function, the element value of the dynamic fusion weight matrix being learned from the correlation strength of the space-time features in different topological states; Step S33: The time-dimension feature vector and the space-dimension feature vector are nonlinearly coupled based on the dynamic fusion weight matrix, wherein the time-dimension feature vector affects the adjacency aggregation mode of the space-dimension feature vector through the state transmission of the gating recurrent unit, so that the weight of the space adjacency matrix is dynamically adjusted over time, and the space-dimension feature vector corrects the time series dependence strength of the time-dimension feature vector through the graph attention coefficient, thereby enhancing the recognition ability of the space-time coupled anomaly, to generate a fused feature vector; Step S34: The fused feature vector is processed through residual connection and layer normalization, the residual connection is used to retain the original space-time feature information, and the layer normalization is used to stabilize the feature distribution in the training process, to generate a line topology feature representation containing space-time joint semantic information; Step S35: Anomaly feature mining is performed on the line topology feature representation, an autoencoder structure is used to reconstruct the line topology feature representation, a reconstruction error is calculated and an error threshold is set, and features with a reconstruction error exceeding the error threshold are determined as abnormal topology features, and a unified topology feature representation containing abnormal topology is generated.

7. The power distribution network oriented spatio-temporal fusion neural network line topology analysis method according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: A dynamic topology state prediction head is constructed, a graph convolution decoder is used to reconstruct the topology relationship of the unified topology feature representation, the convolution kernel parameters of the graph convolution decoder are determined by learning the mapping rule of the topology feature and the connection relationship, and a connection probability matrix between nodes is output, the element value of the connection probability matrix representing the probability of an effective connection between the corresponding nodes; Step S42: A space-time joint loss function is designed, including a time-dimension time series prediction loss and a space-dimension topology structure loss, the time-dimension time series prediction loss being used to constrain the prediction accuracy of the model for the time series change of the electrical quantity, and the space-dimension topology structure loss being used to optimize the prediction accuracy of the node connection relationship, and a dynamic weighting factor being used to balance the optimization objectives of the two dimensions; Step S43: An adaptive learning rate optimization algorithm is used for end-to-end model training, the initial learning rate being set according to the data size and network complexity, the initial learning rate being periodically adjusted through a cosine annealing mechanism in the training process, and an early stopping mechanism being introduced to stop training when the validation set loss does not decrease for a plurality of consecutive rounds, and the model parameters with the optimal performance on the validation set being saved, to construct a dynamic topology state prediction model; Step S44: The real-time collected line operation data is converted into a unified topology feature representation according to the processing procedure of steps S1-S3, input into the trained dynamic topology state prediction model, and a dynamic topology state prediction result is output, including a node connection relationship matrix, a device state health score, and a topology change trend probability distribution. Step S45: Post-processing the dynamic topology state prediction results, and verifying the dynamic topology state prediction results combined with historical topology data to ensure the reliability of the output results, thereby outputting the real-time topology connection relationship and device state change trend corresponding to the power distribution network; based on the real-time topology connection relationship and device state change trend corresponding to the power distribution network, dynamically reconstructing the topology to generate a reliable topology graph at the current time point containing potential branch disconnection and temporary tripping.

8. The power distribution network-oriented spatio-temporal fusion neural network line topology analysis method according to claim 7, characterized in that, The time dimension time series prediction loss in step S42 adopts an improved Huber loss function, and the topology structure loss in the spatial dimension adopts a combination loss mechanism. The specific implementation process includes: In the time dimension loss calculation, the improved Huber loss function realizes differentiated processing of different amplitude electrical quantity fluctuations through a dynamic threshold adjustment module. Squared loss is adopted for small amplitude normal fluctuations to improve prediction accuracy, and absolute value loss is adopted for large amplitude fault fluctuations to enhance the anti-interference ability of the model. The dynamic threshold is adaptively adjusted according to historical electrical quantity fluctuation data and fault statistical rules; In the spatial dimension loss calculation, the cross-entropy loss is used to constrain the classification accuracy of the node connection state, and the prediction accuracy of the node connection relationship of the key tie line and the substation outgoing line is highlighted through a weighting factor, and the influence of the prediction error of the non-key line on the overall topology analysis result is reduced; The link prediction loss is introduced, the Bayesian personalized ranking loss function is used to compare the prediction scores of the actual connection edges and the non-connection edges, and the relative ranking relationship of the edge weights is optimized, so that the model can more accurately identify the effective connection and invalid connection in the topology structure; A loss balancing mechanism is constructed to weight and fuse the time loss and the space loss through dynamic weight adjustment of the weighting factor. The weights of the two are balanced at the beginning of training, and the proportion of the space topology loss is gradually increased with the training process; The loss function is regularized, an L2 regularization term is introduced to suppress the model parameters from being too large, and the Dropout technology is used to randomly shield part of the neurons to prevent model overfitting, thereby designing the corresponding spatio-temporal joint loss function.

9. The power distribution network oriented spatio-temporal fusion neural network line topology analysis method according to claim 7, characterized in that, The post-processing of the dynamic topology state prediction results in step S45 includes: Based on the dynamic topology state prediction results, the basic connectivity indicators of the topology structure are calculated, including the node degree distribution, the average path length and the clustering coefficient; A multi-scenario fault set is generated through Monte Carlo simulation to simulate the influence of different types of faults on the topology structure, and the load transfer margin and power supply recovery time under different fault scenarios are calculated to generate the dynamic response capability indicators of the topology structure; A fault propagation model is constructed to simulate the influence range of a single node fault on the global topology based on the node electrical coupling relationship and the topology connection relationship, and to calculate the fault propagation index to generate the anti-disturbance capability indicators of the topology structure; A topology reliability scoring system is established to weight and fuse the basic connectivity indicators, the dynamic response capability indicators and the anti-disturbance capability indicators to generate the comprehensive reliability score of the topology structure; The potential fault link with a comprehensive reliability score lower than a set threshold is screened out, root cause analysis is performed in combination with historical topology data, electrical characteristics and topology characteristics of the potential fault link are extracted, similarity matching is performed with historical fault cases in the historical topology data, a fault cause is determined, and alarm information including a fault position, possible causes, and an impact range and topology repair suggestions are output; the corresponding fault cause is used to verify and predict the prediction result of the dynamic topology state, so that the corresponding real-time topology connection relationship and device state change trend of the power distribution network are output.

10. The power distribution network-oriented spatio-temporal fusion neural network line topology analysis method according to claim 9, characterized in that, The dynamic topology reconstruction based on the real-time topology connection relationship and device state change trend of the power distribution network in step S45 includes: extracting the real-time topology connection relationship and device state change trend, wherein the real-time topology connection relationship includes connection probability values between nodes, and the device state change trend includes a device health score change curve over time and an abnormal state early warning identifier; a dynamic topology reconstruction rule library is constructed, for a potential branch disconnection scenario, a double determination condition of a connection probability value lower than a disconnection determination threshold and a sudden change of electrical quantity of adjacent nodes is set; for a temporary trip scenario, a joint determination condition of a sudden drop of a device health score, a sudden change of a branch current to zero, and a connection probability of the branch in a topology connection relationship matrix being instantaneously zero is set; topology structure analysis is performed on the real-time topology connection relationship, connection relationships between nodes are traversed based on a connectivity analysis algorithm in graph theory, branches with connection probability values meeting the double determination condition are marked, and branches meeting the joint determination condition are distinguished in combination with the change curve and the abnormal state early warning identifier in the device state change trend, to preliminarily locate fault positions of potential branch disconnection and temporary trip; a multi-source data cross verification mechanism is introduced, the preliminarily located fault positions are compared with switch state signals and fault recording data collected by a real-time monitoring system of the power distribution network, if the connection state of the fault position is consistent with the actual on-off state of the switch and the electrical quantity change matches the fault recording data, the abnormal scenario is confirmed; if there is data deviation, a backtracking verification process is started, the unified topology feature representation generated in step S3 is re-called, and the connection probability values between nodes are corrected; a topology completion algorithm is used to repair the topology structure of the confirmed abnormal scenario, for potential branch disconnection, a virtual backup connection path is supplemented and generated according to the historical connection mode of the disconnected branch and the topology association characteristics of the surrounding nodes; for temporary trip, a temporary replacement branch is marked based on the historical recovery record of the tripped device and the redundant design of the power grid topology, so as to generate a backup path; a visualization generation module of a current time topology graph is constructed, based on device spatial coordinates of a power distribution network geographic information system, different colors are used to identify normal branches, potential disconnected branches, and tripped branches, and the position information of the abnormal branch, the abnormal determination basis, and the recommended treatment scheme including a maintenance priority and a backup path are marked in the current time topology graph. The current time topology graph is verified for reliability, and connectivity, redundancy and load coverage capability indexes of the current time topology graph are calculated. If the connectivity is lower than a power grid safe operation threshold, the redundancy cannot meet the fault transfer requirement or the load coverage capability index has a blind area, the backup path is re-optimized. If all the reliability indexes meet the requirements, the final current time reliable topology graph is output.

Images