Power transmission and transformation system supervision method based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method for monitoring power transmission and transformation systems, comprising: acquiring multi-source operational data of the power transmission and transformation system, preprocessing it, and generating a standard dataset; constructing a dynamic electrical topology graph, correcting the node2vec transition probability, and generating a reliable topology graph; mapping the monitored objects as residual capsules, connecting them to form a capsule array, and generating a residual tensor; coupling the residual tensor and reliability, clustering anomaly clusters, and generating resolution candidates; virtually back-substituting the resolution candidates, recalculating the residual contradiction values ​​of the entire network, and generating a resolution rate; selecting the minimum residual resolution set, and outputting the anomaly source, anomaly type, and monitoring strategy. This invention achieves accurate identification of anomaly sources, anomaly types, and monitoring strategies in power transmission and transformation systems by improving node2vec dynamic topology modeling, electrical causal residual capsule arrays, and minimum residual resolution confirmation.

Description

Technical Field

[0001] This invention relates to the field of intelligent power system monitoring technology, and in particular to a monitoring method for power transmission and transformation systems based on artificial intelligence. Background Technology

[0002] With the expansion of power system scale and the increase in the number of transmission and transformation equipment, the operation and supervision of transmission and transformation systems has gradually evolved from manual inspections, periodic tests, and fixed threshold alarms to intelligent supervision methods relying on SCADA, online monitoring, protection devices, measurement point data collection, and environmental perception data. Existing technologies typically collect data on voltage, current, power, temperature, protection actions, and equipment status, and combine this data with machine learning or anomaly detection models to identify and issue early warnings for the status of transmission lines, transformers, busbars, switchgear, and measurement points.

[0003] However, existing intelligent monitoring methods for power transmission and transformation systems often directly input multi-source data collected into the model for anomaly detection, or construct the electrical topology solely based on static primary wiring relationships and switch remote signaling status. Due to real-world issues such as switch remote signaling errors, abnormal mapping between measurement points and equipment, protection zone configuration deviations, changes in operating modes, and asynchronous communication data, existing methods struggle to accurately determine the reliability of the current electrical topology, easily misinterpreting topology anomalies or measurement point anomalies as equipment malfunctions.

[0004] Existing AI-based monitoring methods typically output a single anomaly probability or risk level, lacking unified constraints on internal electrical conservation, thermoelectric deviations, protection logic, measurement reliability, and silent response relationships. They also lack mechanisms for residual resolution and confirmation of candidate anomalies. When equipment anomalies, measurement point anomalies, topology anomalies, and upstream disturbances are intertwined, existing technologies struggle to accurately determine the anomaly source, anomaly type, and monitored object, leading to a high false alarm rate and impacting the accuracy and enforceability of monitoring conclusions for power transmission and transformation systems.

[0005] Therefore, how to provide an AI-based regulatory approach for power transmission and transformation systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an artificial intelligence-based method for monitoring power transmission and transformation systems. This invention utilizes an improved node2vec model, a dynamic electrical topology reliable mapping, an electrical causal residual capsule array, and a minimum residual resolution-based monitoring confirmation method to distinguish and identify switch remote signaling errors, abnormal measurement point mapping, topology state changes, and equipment anomalies in power transmission and transformation systems. This enables accurate confirmation of anomaly sources, anomaly types, and monitored objects, and has the advantages of low false alarm rate, accurate anomaly location, strong interpretability of monitoring conclusions, and high engineering applicability.

[0007] The power transmission and transformation system monitoring method based on artificial intelligence according to embodiments of the present invention includes:

[0008] Acquire multi-source operation data from the power transmission and transformation system, preprocess the multi-source operation data, and obtain a standard dataset;

[0009] A dynamic electrical topology graph is constructed based on a standard dataset. The power flow direction vector, voltage phase angle difference, and switch state confidence extracted from the standard dataset within a preset refresh cycle are used as transfer weight factors to modify the random walk transfer probability function in the improved node2vec model. The node embedding update is performed on the dynamic electrical topology graph through the improved node2vec model, and the node confidence and edge confidence are calculated to obtain a reliable dynamic electrical topology graph.

[0010] Based on the dynamic electrical topology credibility graph, each regulated object is mapped as an electrical causal residual capsule. Each electrical causal residual capsule is connected according to the node credibility and edge credibility to form an electrical causal residual capsule array, generating multiple types of residual tensors.

[0011] Couple multiple types of residual tensors, node credibility and edge credibility, perform embedded spatial local density clustering on dynamic electrical topology credibility graph to identify anomalous clusters, and use residual minimum cut entropy reduction strategy to generate node-anomaly type binary set for each anomalous cluster to obtain interpretable resolution candidates.

[0012] The explainable and resolvable candidate options are virtually substituted back into the dynamic electrical topology credibility graph and the electrical causal residual capsule array, and the residual contradiction value of the whole network after the substitution is recalculated to generate the residual resolution rate of the corresponding candidate options;

[0013] The minimum residual resolution set is determined based on the residual resolution rate, topological constraints, protection logic constraints, and quiescent constraints. Based on the minimum residual resolution set, the abnormal source, abnormal type, abnormal impact range, and corresponding regulatory strategy in the power transmission and transformation system are output.

[0014] Optionally, the multi-source operating data includes voltage data, current data, active power data, reactive power data, voltage phase angle data, frequency data, switch status data, protection action data, equipment temperature data, partial discharge data, gas pressure data, surge arrester leakage current data, circuit breaker mechanical characteristic data, measuring point number data, measuring point binding relationship data, communication status data, ambient temperature data, humidity data, wind speed data, primary wiring data, equipment connection relationship data, and protection zone relationship data.

[0015] Optionally, the preprocessing of multi-source operating data to obtain a standard dataset includes: uniformly timestamping the multi-source operating data; aligning the data according to a preset sampling period; removing outliers that exceed the rated operating range of the equipment; completing missing data; unifying the mapping relationship between measurement point numbers and equipment numbers; converting data from different sources into a unified data format; and normalizing voltage data, current data, active power data, reactive power data, voltage phase angle data, equipment status data, and environmental data to obtain a standard dataset.

[0016] Optionally, obtaining the dynamic electrical topology reliability graph includes:

[0017] Based on the primary wiring data, equipment connection relationships, switch status data, measurement point binding relationships, and protection zone relationships in the standard dataset, a dynamic electrical topology diagram containing equipment nodes, measurement point nodes, protection nodes, and electrical connection edges is constructed, and the effective status of each electrical connection edge is determined according to the switch opening and closing status.

[0018] Within a preset refresh cycle, the power flow direction vector, voltage phase angle difference, and switch status confidence level corresponding to each valid electrical connection edge are extracted from the standard dataset. The power flow direction vector is converted into power flow direction consistency weight, the voltage phase angle difference is converted into phase angle consistency weight, and the switch status confidence level is converted into switch confidence weight.

[0019] For a random walk process of the improved node2vec model, the current walking node, the previous walking node, and the candidate transition nodes of the current walking node are determined. The candidate transition nodes are the adjacent nodes that are connected to the current walking node through an effective electrical connection edge.

[0020] Based on the return and input / output parameters of the node2vec model, the structural transfer weight of each candidate transfer node relative to the previous roaming node is determined. The structural transfer weight, power flow direction consistency weight, phase angle consistency weight, and switch confidence weight are then fused to obtain the comprehensive transfer weight of each candidate transfer node.

[0021] The combined transfer weights of all candidate transfer nodes corresponding to the same current walk node are summed, and the ratio of the combined transfer weight of each candidate transfer node to the summation result is determined as the random walk transfer probability of the candidate transfer node. The random walk transfer probabilities of each candidate transfer node constitute the random walk transfer probability function of the improved node2vec model.

[0022] The node walking sequence of the dynamic electrical topology graph is generated based on the random walk transition probability function. The node embedding sequence is updated by improving the node2vec model to obtain the node embedding vector of each node. The node credibility and edge credibility are calculated based on the node embedding vector, historical normal embedding vector, effective state of electrical connection edge and time-varying edge weight to obtain the dynamic electrical topology credibility graph.

[0023] Optionally, the formation of the electrical causal residual capsule array to generate multiple types of residual tensors includes:

[0024] Based on the dynamic electrical topology credibility graph, the regulatory objects in the power transmission and transformation system are determined. The busbar, line, transformer, circuit breaker, disconnector, cable joint, bushing, surge arrester, protection device and measuring point are treated as independent regulatory objects. A corresponding electrical causal residual capsule is created for each regulatory object.

[0025] The port configuration of the corresponding electrical causal residual capsule is determined according to the equipment type of the regulated object. The electrical causal residual capsule corresponding to the busbar is configured as an incoming port, an outgoing port, a bus tie port, a voltage phase angle port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the line is configured as a head electrical port, an end electrical port, a temperature port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the transformer is configured as a high-voltage side port, a medium-voltage side port, a low-voltage side port, a tap changer port, a temperature port, a cooling status port, a measurement port, and a protection port.

[0026] The voltage, current, active power, reactive power, voltage phase angle, equipment temperature, protection action, measurement point, communication status, and environmental data that are consistent with the regulated object identifier in the standard dataset are connected to the port of the corresponding electrical causal residual capsule. The node credibility and edge credibility in the dynamic electrical topology credibility graph are written into the corresponding electrical causal residual capsule.

[0027] According to the effective connection relationship in the dynamic electrical topology confidence graph, each electrical causal residual capsule is connected. When there is an effective electrical connection edge between two monitored objects and the confidence of the corresponding edge meets the preset connection conditions, a residual propagation connection is established between the two electrical causal residual capsules. When there is a binding relationship between the monitored object and the measuring point or protection device, a measurement association connection or protection association connection is established between the corresponding electrical causal residual capsules to form an electrical causal residual capsule array.

[0028] By generating conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals for each regulated object through an electrical causal residual capsule array, and arranging them according to the regulated object, residual type, and sampling time to generate multi-type residual tensors.

[0029] Optionally, obtaining the interpretable resolution candidate includes:

[0030] We obtain multiple types of residual tensors, node credibility, edge credibility, and node embedding results obtained by the improved node2vec model during the node embedding update process. We then concatenate the residual type, residual magnitude, node credibility, associated edge credibility, and node embedding results corresponding to the same regulatory object to form the clustering features of the regulatory object.

[0031] In the dynamic electrical topology credibility graph, the cluster feature distance of the regulated objects is used as the embedding space distance, and the edge credibility is used as the topological connection constraint to determine the effective neighborhood of each regulated object. The local density of the regulated objects is determined based on the number of residual similar regulated objects in the effective neighborhood.

[0032] The regulatory objects whose local density reaches the first density threshold and whose multi-type residual tensors reach the first residual threshold are identified as cluster cores. The regulatory objects that are within the embedding space distance of the cluster cores and are connected by trusted edges are merged into the same cluster to obtain the abnormal cluster.

[0033] For each abnormal cluster, candidate splitting edges are determined based on the residual differences between adjacent monitored objects within the cluster, edge credibility, and residual propagation direction. The distribution changes of conserved residuals, thermoelectric deviation residuals, protection logic residuals, measurement credibility residuals, silent violation residuals, and propagation residuals in each sub-cluster before and after splitting are statistically analyzed. The candidate splitting edge whose residual distribution entropy decreases to the first entropy reduction threshold and has the lowest splitting cost is determined as the target splitting edge.

[0034] The abnormal clusters are segmented according to the target segmentation edge. The regulatory objects with the highest residual concentration after segmentation and whose node credibility and edge credibility meet the abnormal location conditions are identified as abnormal candidate nodes. The abnormal type is determined according to the dominant residual type corresponding to the abnormal candidate node. The abnormal candidate node and the abnormal type are combined to generate a set of node-abnormal type binary pairs to obtain the interpretable resolution candidate.

[0035] Optionally, the generation of the residual resolution rate of the corresponding candidate options includes:

[0036] Read the set of node-anomaly type pairs in the interpretable resolution candidate, and determine the regulatory object to be back-submitted, the anomaly type to be back-submitted, and the location of the back-submitted action based on each node-anomaly type pair.

[0037] Virtual back-substitution processing is performed according to the type of anomaly to be back-substituted. When the anomaly to be back-substituted is a measurement point anomaly, measurement isolation or substitution input processing is performed on the corresponding measurement port. When the anomaly to be back-substituted is a topology state anomaly, the effective state and edge confidence of the corresponding node or connection edge are corrected. When the anomaly to be back-substituted is a device body anomaly, a body anomaly mark is written to the corresponding electrical causal residual capsule and its conservation ledger port or thermoelectric deviation port is updated. When the anomaly to be back-substituted is an upstream disturbance, a disturbance input is written to the corresponding boundary node. When the anomaly to be back-substituted is a protection communication anomaly, protection signal isolation or communication state correction processing is performed on the corresponding protection latching port.

[0038] After completing the virtual back-substitution process, the electrical causal residual capsule array is reconnected based on the back-substituted dynamic electrical topology credibility graph. The conserved residual, thermoelectric deviation residual, protection logic residual, measurement credibility residual, silent violation residual, and propagation residual are regenerated through the reconnected electrical causal residual capsule array.

[0039] The total amount of conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, propagation residuals, and topological reliability residuals determined by node reliability and edge reliability are calculated separately after back-substitution. The total amount of each residual is weighted and summed according to the preset weights to obtain the total residual contradiction value of the whole network after back-substitution.

[0040] The difference between the total network residual contradiction value before virtual back-substitution and the total network residual contradiction value after back-substitution is calculated. The ratio of the difference to the total network residual contradiction value before virtual back-substitution is calculated to obtain the residual resolution rate of the corresponding interpretable resolution candidate.

[0041] Optionally, the step of outputting the anomaly source, anomaly type, anomaly impact range, and corresponding regulatory strategy in the power transmission and transformation system based on the minimum residual resolution set includes:

[0042] Read the residual resolution rate corresponding to each explainable resolution candidate, filter the explainable resolution candidates according to the residual resolution rate threshold, and retain the candidates whose residual resolution rate reaches the residual resolution rate threshold as the candidate resolution options;

[0043] A topology constraint verification is performed on the candidate to be eliminated. If the candidate does not generate new invalid electrical connections, does not destroy the valid connection relationship in the dynamic electrical topology confidence graph, and does not make the edge confidence lower than the preset edge confidence threshold after the candidate is substituted back, the candidate is determined to meet the topology constraint conditions.

[0044] The candidate elimination option is checked for protection logic constraints and silent constraints. When the protection action status after the candidate option is replaced is consistent with the protection closing port of the corresponding supervised object, and no abnormal response is found in the node in the silent state in the dynamic electrical topology confidence diagram, the candidate option is determined to meet the protection logic constraints and silent constraints.

[0045] Candidates that simultaneously satisfy the residual resolution rate threshold, topological constraints, protection logic constraints, and quiescent constraints are combined and screened in order of increasing number of candidates and increasing regulatory cost. The combination of candidates with the fewest number of candidates and the lowest regulatory cost is determined as the minimum residual resolution set.

[0046] The anomaly source and anomaly type are determined based on the node-anomaly type tuple in the minimum residual resolution set. The anomaly impact range is determined based on the trusted connection path of the anomaly source in the dynamic electrical topology trusted graph. The corresponding regulatory strategy is generated based on the anomaly type.

[0047] The beneficial effects of this invention are:

[0048] This invention constructs a dynamic electrical topology reliability graph by improving the node2vec model. It can combine power flow direction, voltage phase angle difference and switch status reliability to correct the node embedding process, so that the supervision of power transmission and transformation system no longer relies solely on static primary wiring relationship or switch remote signaling status. This improves the reliability of current electrical topology identification and reduces misjudgments caused by remote signaling errors, abnormal measurement point binding or changes in operation mode.

[0049] This invention generates multiple types of residual tensors through an electrical causal residual capsule array, which can uniformly express the conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals of various monitored objects. This makes the differences between equipment body anomalies, measurement point anomalies, topology anomalies, and protection communication anomalies clearer, and improves the accuracy and interpretability of anomaly identification.

[0050] This invention employs a minimum residual resolution-based regulatory confirmation method. By performing virtual back substitution on interpretable resolution candidates and calculating the residual resolution rate, it can determine the minimum residual resolution set from multiple anomaly candidates, accurately output the anomaly source, anomaly type, anomaly impact range, and regulatory strategy, reduce equipment false alarms in intelligent monitoring of power transmission and transformation systems, and improve the pertinence and engineering feasibility of regulatory handling. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1This is a flowchart of the artificial intelligence-based power transmission and transformation system monitoring method proposed in this invention;

[0053] Figure 2 This is a schematic diagram illustrating the process of constructing a dynamic electrical topology credibility graph based on an improved node2vec model in the power transmission and transformation system monitoring method based on artificial intelligence proposed in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figure 1 and Figure 2 Artificial intelligence-based methods for regulating power transmission and transformation systems include:

[0056] Acquire multi-source operation data from the power transmission and transformation system, preprocess the multi-source operation data, and obtain a standard dataset;

[0057] A dynamic electrical topology graph is constructed based on a standard dataset. The power flow direction vector, voltage phase angle difference, and switch state confidence extracted from the standard dataset within a preset refresh cycle are used as transfer weight factors to modify the random walk transfer probability function in the improved node2vec model. The node embedding update is performed on the dynamic electrical topology graph through the improved node2vec model, and the node confidence and edge confidence are calculated to obtain a reliable dynamic electrical topology graph.

[0058] Based on the dynamic electrical topology credibility graph, each regulated object is mapped as an electrical causal residual capsule. Each electrical causal residual capsule is connected according to the node credibility and edge credibility to form an electrical causal residual capsule array, generating multiple types of residual tensors.

[0059] Couple multiple types of residual tensors, node credibility and edge credibility, perform embedded spatial local density clustering on dynamic electrical topology credibility graph to identify anomalous clusters, and use residual minimum cut entropy reduction strategy to generate node-anomaly type binary set for each anomalous cluster to obtain interpretable resolution candidates.

[0060] The explainable and resolvable candidate options are virtually substituted back into the dynamic electrical topology credibility graph and the electrical causal residual capsule array, and the residual contradiction value of the whole network after the substitution is recalculated to generate the residual resolution rate of the corresponding candidate options;

[0061] The minimum residual resolution set is determined based on the residual resolution rate, topological constraints, protection logic constraints, and quiescent constraints. Based on the minimum residual resolution set, the abnormal source, abnormal type, abnormal impact range, and corresponding regulatory strategy in the power transmission and transformation system are output.

[0062] In this embodiment, the multi-source operating data includes voltage data, current data, active power data, reactive power data, voltage phase angle data, frequency data, switch status data, protection action data, equipment temperature data, partial discharge data, gas pressure data, surge arrester leakage current data, circuit breaker mechanical characteristic data, measuring point number data, measuring point binding relationship data, communication status data, ambient temperature data, humidity data, wind speed data, primary wiring data, equipment connection relationship data, and protection zone relationship data.

[0063] In this embodiment, the preprocessing of multi-source operating data to obtain a standard dataset includes: uniformly timestamping the multi-source operating data; aligning the time according to a preset sampling period; removing outliers that exceed the rated operating range of the equipment; completing missing data; unifying the mapping relationship between measurement point numbers and equipment numbers; converting data from different sources into a unified data format; and normalizing voltage data, current data, active power data, reactive power data, voltage phase angle data, equipment status data, and environmental data to obtain a standard dataset.

[0064] In this embodiment, obtaining the dynamic electrical topology reliability map includes:

[0065] Based on the primary wiring data, equipment connection relationships, switch status data, measurement point binding relationships, and protection zone relationships in the standard dataset, a dynamic electrical topology diagram containing equipment nodes, measurement point nodes, protection nodes, and electrical connection edges is constructed. The effective state of each electrical connection edge is determined according to the switch's open / closed state. Specifically, the construction of the dynamic electrical topology diagram containing equipment nodes, measurement point nodes, protection nodes, and electrical connection edges is as follows:

[0066] Extract busbars, lines, transformers, circuit breakers, disconnectors, and unique identifiers from the primary wiring data, and establish equipment nodes according to equipment type;

[0067] The binding relationship of measurement points is analyzed, and voltage measurement points, current measurement points, power measurement points, and temperature measurement points are associated with their corresponding equipment identifiers to establish measurement point nodes;

[0068] Read the protection zone relationship, associate the differential protection device, distance protection device, bus differential protection device and the equipment identifiers corresponding to the protection zone, and establish protection nodes;

[0069] Based on the equipment connection relationship, establish electrical connection edges between the busbar and the equipment node, and between the equipment nodes, and mark the current open or closed status of the connection edges according to the switch status data;

[0070] Electrical connection edges in the open state are set as invalid edges, and electrical connection edges in the closed state are set as valid edges, forming a dynamic electrical topology diagram that includes equipment nodes, measurement point nodes, protection nodes, and valid electrical connection edges.

[0071] Within a preset refresh cycle, the power flow direction vector, voltage phase angle difference, and switch state confidence level corresponding to each valid electrical connection edge are extracted from the standard dataset. The power flow direction vector is converted into a power flow direction consistency weight, the voltage phase angle difference is converted into a phase angle consistency weight, and the switch state confidence level is converted into a switch confidence weight, wherein:

[0072] The power flow direction vector is converted into a power flow direction consistency weight. Specifically, for each valid electrical connection edge, the active power and current direction at both ends of the edge are read, and the specified power flow direction is determined in combination with the primary wiring. The real-time power flow direction is compared with the specified power flow direction. If the comparison results are consistent and the continuous consistency duration is not less than two sampling periods, it is recorded as completely consistent and assigned a power flow direction consistency weight of 1. If the directions are opposite and the continuous opposite duration exceeds two sampling periods, it is recorded as completely inconsistent and assigned a power flow direction consistency weight of zero. If the direction changes alternately within two sampling periods, the power flow direction consistency weight is calculated according to the proportion of the consistency duration and rounded to three decimal places.

[0073] The voltage phase angle difference is converted into a phase angle consistency weight. Specifically, the voltage phase angles measured synchronously at both ends of the side are read, and the absolute value of the real-time voltage phase angle difference is calculated. If the phase angle difference is no greater than five degrees, the phase angle consistency weight is assigned as one. If the phase angle difference is greater than five degrees but not greater than fifteen degrees, the phase angle consistency weight is linearly calculated by dividing (fifteen minus the current difference) by ten. If the phase angle difference is greater than fifteen degrees, the phase angle consistency weight is assigned as zero.

[0074] The switch status reliability is converted into a switch reliability weight. Specifically, the switch status reliability identifier is read, which includes three indicators: remote signaling integrity, remote signaling and power flow direction matching degree, and remote signaling jitter count. When all three indicators meet the requirements of remote signaling integrity, power flow matching, and jitter count not exceeding one, the switch reliability weight is assigned as 1. If two of them are met, the switch reliability weight is assigned as 0.6. If only one of them is met, the switch reliability weight is assigned as 0.3. If none of them are met, the switch reliability weight is assigned as 0.00.

[0075] For a random walk process of the improved node2vec model, the current walking node, the previous walking node, and the candidate transition nodes of the current walking node are determined. The candidate transition nodes are the adjacent nodes that are connected to the current walking node through an effective electrical connection edge.

[0076] Based on the return and input / output parameters of the node2vec model, the structural transfer weight of each candidate transfer node relative to the previous roaming node is determined. The structural transfer weight, power flow direction consistency weight, phase angle consistency weight, and switch reliability weight are then fused to obtain the comprehensive transfer weight of each candidate transfer node. Specifically, the structural transfer weight of each candidate transfer node relative to the previous roaming node is determined as follows:

[0077] Determine the topological relationship between the candidate transfer node and the previous traversing node: if the two nodes are the same, they are classified as the return category; if the candidate transfer node and the previous traversing node are directly connected by a valid electrical connection edge, they are classified as the inward category; if the two nodes are not directly connected and the shortest reliable path length between the candidate transfer node and the previous traversing node is greater than one, they are classified as the outward category.

[0078] The highest structure transfer weight is assigned to the return category, with a weight of one; the medium structure transfer weight is assigned to the inward category, with a weight equal to one multiplied by the return bias factor; and the lowest structure transfer weight is assigned to the outward category, with a weight equal to one multiplied by the inward bias factor.

[0079] The structural transfer weight obtained for each candidate transfer node is multiplied by the power flow direction consistency weight, phase angle consistency weight, and switch reliability weight to obtain the comprehensive transfer weight of the candidate transfer node.

[0080] The overall transfer weights of all candidate transfer nodes corresponding to the same current walking node are summed, and the random walk transfer probability of the candidate transfer node is determined by the proportion of each overall transfer weight to the total weight.

[0081] Write the probability distribution consisting of the random walk transition probabilities of all candidate transition nodes into the improved node2vec model to complete the update of the random walk transition probability function for the current step.

[0082] The combined transition weights of all candidate transition nodes corresponding to the same current walking node are summed. The ratio of the combined transition weight of each candidate transition node to the summation result is determined as the random walk transition probability of the candidate transition node. The random walk transition probabilities of each candidate transition node constitute the random walk transition probability function of the improved node2vec model, specifically as follows:

[0083] Allocate a transition probability record area for the current wandering node, and write the node number - random walk transition probability tuple according to the order of the candidate transition nodes in the adjacency list;

[0084] During the writing process, the transition probability is prefixed and accumulated, and each tuple is expanded into a triplet of node number - start interval - end interval, where the start interval is the end interval of the previous node, and the end interval is the start interval plus the random walk transition probability of the current node.

[0085] Store all triples sequentially into a probability mapping table indexed by the current wandering node number;

[0086] During random walk, by generating a uniform random number between zero and one, querying the probability mapping table for triples containing that random number, the next walk node number can be directly obtained, realizing random sampling of adjacent nodes based on comprehensive transfer weights;

[0087] The above construction and query mechanism is repeated for each node in the graph, and finally the random walk transition probability function of the improved node2vec model is completed in the form of a node number-probability interval list.

[0088] The node walk sequence of the dynamic electrical topology graph is generated based on the random walk transition probability function. The node embedding sequence is updated using an improved node2vec model to obtain the node embedding vector for each node. Based on the node embedding vector, historical normal embedding vector, effective state of electrical connection edges, and time-varying edge weights, the node credibility and edge credibility are calculated to obtain the dynamic electrical topology credibility graph. Specifically, the calculation of node credibility and edge credibility to obtain the dynamic electrical topology credibility graph is as follows:

[0089] For each node in the graph, call the latest embedding vector generated by the improved node2vec model, and read the baseline embedding vector of the node in its historical normal operating state;

[0090] First, multiply the elements and sum them to get the dot product of the current vector and the reference vector. Then, calculate the square root of the sum of the squares of the elements of the two vectors as the vector magnitude. Divide the dot product value by the product of the magnitudes of the two vectors to get the similarity between zero and one. Write the similarity into the node attribute as the node credibility.

[0091] Traverse each electrical connection edge that is in a valid state, read the node confidence of the two ends of the connection edge and the time-varying edge weight of the connection edge, average the node confidence of the two ends and multiply it by the time-varying edge weight to obtain the comprehensive confidence value of the connection edge, and write the comprehensive confidence value into the edge attribute as the edge confidence.

[0092] The node confidence level is compared with a preset node confidence level threshold, and nodes with confidence levels below the threshold are marked as low-confidence nodes. The edge confidence level is compared with a preset edge confidence level threshold, and edges with confidence levels below the threshold are marked as low-confidence edges. Low-confidence edges are also considered invalid electrical connections and are removed from the topology analysis.

[0093] A dynamic electrical topology credibility graph is generated using all nodes and their credibility, and all valid connected edges and their credibility as elements.

[0094] In this embodiment, the formation of an electrical causal residual capsule array to generate multiple types of residual tensors includes:

[0095] Based on the dynamic electrical topology reliability graph, the monitored objects in the power transmission and transformation system are determined. Busbars, lines, transformers, circuit breakers, disconnectors, cable joints, bushings, surge arresters, protection devices, and measuring points are each treated as independent monitored objects. A corresponding electrical causal residual capsule is created for each monitored object. Specifically, the creation of the corresponding electrical causal residual capsule for each monitored object is as follows:

[0096] Assign a unique capsule number to the current regulated object and write the node number, the corresponding switch bay, voltage level and protection zone in the dynamic electrical topology trust graph into the capsule metadata area;

[0097] Based on the type of regulated object, load the preset port templates: If it is a busbar, create an incoming port list, an outgoing port list, a bus tie port, a phase angle monitoring port, and a protection port; if it is a line, create a head-end voltage port, a head-end current port, a tail-end voltage port, a tail-end current port, a temperature port, and a protection port; if it is a transformer, create a high-voltage side port, a medium-voltage side port, a low-voltage side port, a tap changer port, an oil temperature port, and a differential protection port; if it is a circuit breaker or disconnector, create upstream and downstream current ports, contact temperature ports, mechanism status ports, and switch protection ports, and load the corresponding templates for other objects in sequence.

[0098] Write the rated capacity, impedance parameters, and initial loss coefficient into the conservation ledger area of ​​the capsule, set all ledger fields to zero, and record the sampling timestamp;

[0099] Write the ambient temperature channel, equipment temperature rise curve channel, and cooling status channel into the thermoelectric deviation zone, and initially set the deviation value to zero.

[0100] In the protection engagement area, establish a protection action register, a protection start register, and a protection lockout register, and set them all to the inactive state;

[0101] Record the list of associated measurement point numbers, communication status register, and measurement point integrity flag in the measurement isolation area, and set the measurement point connectivity status to normal.

[0102] Write the list of silent ports determined by the current topology trust graph in the silent constraint area, and clear the silent violation counter to zero;

[0103] In the residual propagation region, initialize the upstream residual buffer, downstream residual buffer, and lateral residual buffer, set the residual values ​​to zero, and write the propagation weights with adjacent capsules.

[0104] After completing port establishment and initialization of each functional area, the capsule object is written into the electrical causal residual capsule array, using the capsule number-node sequence number-object type triplet as the index;

[0105] The port configuration of the corresponding electrical causal residual capsule is determined according to the equipment type of the regulated object. The electrical causal residual capsule corresponding to the busbar is configured as an incoming port, an outgoing port, a bus tie port, a voltage phase angle port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the line is configured as a head electrical port, an end electrical port, a temperature port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the transformer is configured as a high-voltage side port, a medium-voltage side port, a low-voltage side port, a tap changer port, a temperature port, a cooling status port, a measurement port, and a protection port.

[0106] The voltage, current, active power, reactive power, voltage phase angle, equipment temperature, protection action, measurement point, communication status, and environmental data that are consistent with the regulated object identifier in the standard dataset are connected to the port of the corresponding electrical causal residual capsule. The node credibility and edge credibility in the dynamic electrical topology credibility graph are written into the corresponding electrical causal residual capsule.

[0107] According to the effective connection relationship in the dynamic electrical topology confidence graph, each electrical causal residual capsule is connected. When there is an effective electrical connection edge between two monitored objects and the corresponding edge confidence meets the preset connection condition, a residual propagation connection is established between the two electrical causal residual capsules. When there is a binding relationship between the monitored object and the measuring point or protection device, a measurement association connection or protection association connection is established between the corresponding electrical causal residual capsules to form an electrical causal residual capsule array. The preset connection condition is that when there is an effective electrical connection edge between two monitored objects and the edge confidence of the effective electrical connection edge is not lower than the preset edge confidence threshold, and the node confidence of the nodes at both ends of the effective electrical connection edge is not lower than the preset node confidence threshold, the preset connection condition is determined to be met.

[0108] The electrical causal residual capsule array is formed as follows:

[0109] Traverse each valid electrical connection edge in the dynamic electrical topology confidence graph, extract the node numbers at both ends of the connection edge, and retrieve the corresponding electrical causal residual capsules respectively; if the edge confidence of the connection edge reaches the preset edge confidence threshold and the node confidence of both ends of the node reaches the preset node confidence threshold, then create a new propagation channel record in the residual propagation area of ​​each of the two capsules, write the peer capsule number, the connection type code as electrical propagation, the propagation weight as edge confidence, and register the bidirectional propagation relationship in the capsule index table;

[0110] For the measurement point binding relationship, a one-to-one measurement association is established between the measurement point capsule corresponding to the measurement point and the device capsule to which it belongs. The association channel records are written to each other in the measurement isolation area between the two, the connection type code is recorded as the measurement association, the communication reliability is written into the channel weight, and the measurement association relationship is registered in the capsule index table.

[0111] To address the binding relationship between the protection device and the protected equipment, a one-to-many protection association is established between the protection capsule and each protected equipment capsule. The association channel records are written to each other in the protection engagement area of ​​the two, the connection type code is recorded as the protection association, the protection coverage weight is written to the channel weight, and the protection association relationship is registered in the capsule index table.

[0112] After completing the writing of the three types of channels—electrical propagation, measurement association, and protection association—the capsule array initialization routine is executed. All capsule objects are loaded into a contiguous memory area in order of capsule number, generating a capsule pointer array. Then, according to the capsule index table, an independent adjacency list is established for each connection type, and the corresponding capsule number pairs are written into the list nodes and linked into the adjacency list.

[0113] The equipment capsule, measuring point capsule, and protection capsule are interconnected through three logical channels: the electrical propagation chain, the measurement association chain, and the protection association chain, forming an electrical causal residual capsule array;

[0114] The conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals for each regulated object are generated using an electrical causality residual capsule array. These residual tensors are then arranged according to the regulated object, residual type, and sampling time to generate multiple types of residual tensors. Specifically, the conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals for each regulated object are generated as follows:

[0115] Conservation residual: At each sampling time, the active power and reactive power of all electrical ports of the capsule are read. The sum of the power flowing into the port is compared with the sum of the power flowing out of the port and the preset loss value. The absolute value of the difference is the conservation residual. If the monitored object is a transformer, the copper loss and iron loss calculated according to the impedance parameters are deducted, and the same difference comparison is performed again.

[0116] Thermoelectric deviation residual: Read the load current, ambient temperature and equipment temperature rise curve of the capsule, look up the expected temperature rise according to the current load and environmental conditions, and calculate the difference between the expected temperature rise and the real-time temperature measurement port reading. The absolute value of the difference is taken as the thermoelectric deviation residual.

[0117] Protection logic residual: Read the protection action register and the body residual signal of the capsule. If the body residual signal reaches the protection activation / deactivation setting value but the corresponding protection action register is not set, or the protection action register is set but the body residual signal is lower than the protection activation / deactivation setting value, it is recorded as a protection logic inconsistency. The protection logic residual is obtained by accumulating the number of inconsistencies.

[0118] Measurement Reliability Residual: For each measurement port, compare its communication status, continuous sampling stability, and corresponding measurement values ​​of adjacent capsules. When communication is abnormal, sampling jitter exceeds twice, or the deviation from the measurement value of adjacent capsules exceeds the limit, record the measurement abnormality count and write the count value into the measurement reliability residual.

[0119] Silent violation residual: Check the list of silent constraint ports. If a silent port experiences a sudden change in its electrical quantity, protection action, or measurement abnormality in the current sampling period, it is recorded as a silent violation. The number of silent violations is the silent violation residual.

[0120] Propagation residual: Read the residual signals of the same type in the propagation port of the capsule and the propagation port of the adjacent capsule. If the propagation direction is expected to be consistent but there is an opposite direction, a sudden change in amplitude or an abnormal attenuation rate, then the number of such abnormalities is accumulated, and the accumulated value is the propagation residual.

[0121] In this embodiment, obtaining the explainable resolution candidate includes:

[0122] We obtain multiple types of residual tensors, node credibility, edge credibility, and node embedding results obtained by the improved node2vec model during the node embedding update process. We then concatenate the residual type, residual magnitude, node credibility, associated edge credibility, and node embedding results corresponding to the same regulatory object to form the clustering features of the regulatory object.

[0123] In the dynamic electrical topology credibility graph, the cluster feature distance of the regulated objects is used as the embedding space distance, and the edge credibility is used as the topological connection constraint to determine the effective neighborhood of each regulated object. The local density of the regulated object is determined based on the number of residual similar regulated objects within the effective neighborhood. Specifically, determining the effective neighborhood of each regulated object and determining the local density of the regulated object based on the number of residual similar regulated objects within the effective neighborhood is as follows:

[0124] For each monitored object, first read the node embedding vector, node confidence, and residual vector corresponding to the multi-type residual tensor at the current sampling time. Then, concatenate the node embedding vector and the residual vector in a preset order to obtain the clustering feature vector.

[0125] Traverse the adjacent nodes in the dynamic electrical topology confidence graph that are connected to the monitored object. If the edge confidence of the connecting edge is not lower than the edge confidence threshold and the node confidence of the adjacent node is not lower than the node confidence threshold, then the adjacent node is determined to be a candidate neighbor.

[0126] Calculate the Euclidean distance between the cluster feature vector of the candidate neighbor and the cluster feature vector of the current monitored object. If the distance is not greater than the preset embedding distance threshold, add the candidate neighbor to the effective neighborhood of the current monitored object and record the edge confidence as the neighborhood weight.

[0127] Continue iterating through all candidate neighbors until no new candidate neighbors are available. Write the node numbers of all neighbor nodes that meet the conditions into the list of valid neighborhoods, and write the corresponding neighborhood weights at the same time.

[0128] Read the residual vector of each neighbor node in the effective neighborhood list, calculate the absolute difference with the residual vector of the current monitored object, and if the difference is not greater than the residual similarity threshold, it is counted as a residual similar neighbor. The neighborhood weight corresponding to the neighbor is accumulated into the local density counter.

[0129] After the traversal is complete, the accumulated value in the local density counter is the local density of the current monitored object;

[0130] The regulatory objects whose local density reaches the first density threshold and whose multi-type residual tensors reach the first residual threshold are identified as cluster cores. The regulatory objects that are within the embedding space distance of the cluster cores and are connected by trusted edges are merged into the same cluster to obtain the abnormal cluster.

[0131] For each abnormal cluster, candidate splitting edges are determined based on the residual differences between adjacent monitored objects within the cluster, edge credibility, and residual propagation direction. The distribution changes of conserved residuals, thermoelectric deviation residuals, protection logic residuals, measurement credibility residuals, silent violation residuals, and propagation residuals in each sub-cluster before and after splitting are statistically analyzed. The candidate splitting edge whose residual distribution entropy decreases to the first entropy reduction threshold and has the lowest splitting cost is determined as the target splitting edge.

[0132] The abnormal clusters are segmented according to the target segmentation edges. The monitored objects with the highest residual concentration after segmentation and whose node and edge credibility meet the anomaly location criteria are identified as candidate anomaly nodes. The anomaly type is determined based on the dominant residual type corresponding to the candidate anomaly node. The candidate anomaly nodes and anomaly types are combined to generate a set of node-anomaly type tuples, yielding interpretable resolution candidates. Specifically, the interpretable resolution candidates are as follows:

[0133] After the target cutting edge is determined, the connection of the abnormal cluster is broken along each target cutting edge, and the cluster is split into several sub-clusters;

[0134] For each sub-cluster, the sum of the various types of residuals of the internally regulated objects is calculated, and the ratio of the sum of the sub-cluster residuals to the number of regulated objects in that sub-cluster is calculated. The sub-cluster with the highest ratio is taken as the sub-cluster with the highest residual concentration.

[0135] In the sub-cluster with the highest residual concentration, the node credibility and the credibility of the directly connected edges are read for each monitored object. If the node credibility is higher than the node credibility threshold and the edge credibility of all connected edges is higher than the edge credibility threshold, the monitored object is retained as an anomaly location candidate.

[0136] For the remaining candidate regulatory targets, compare the numerical values ​​of conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals, and mark the residual type with the largest value as the dominant residual type of the regulatory target;

[0137] Generate binary entries according to the format of regulatory object number - dominant residual type, and write the entries corresponding to all candidate regulatory objects into the node-anomaly type binary set.

[0138] In this embodiment, the generation of the residual elimination rate of the corresponding candidate options includes:

[0139] Read the set of node-anomaly type pairs from the interpretable resolution candidates. Based on each node-anomaly type pair, determine the object to be back-upgraded for monitoring, the anomaly type to be back-upgraded for monitoring, and the location of the action to be back-upgraded for monitoring. Specifically, determining the object to be back-upgraded for monitoring, the anomaly type to be back-upgraded for monitoring, and the location of the action to be back-upgraded for monitoring is as follows:

[0140] Read the node-anomaly type binary entries one by one. First, retrieve the corresponding electrical causal residual capsule pointer in the capsule index table by node number to determine the type of the regulated object to which the capsule belongs and the node number in the dynamic electrical topology confidence graph. Then, write the retrieved capsule pointer into the list of regulated objects to be returned.

[0141] Parse the anomaly type field in the binary tuple. If the anomaly type is measurement point anomaly, mark the anomaly type to be back-replaced as measurement point anomaly. If the anomaly type is topology status anomaly, mark the anomaly type to be back-replaced as topology status anomaly. If the anomaly type is equipment body anomaly, mark the anomaly type to be back-replaced as equipment body anomaly. If the anomaly type is upstream disturbance anomaly, mark the anomaly type to be back-replaced as upstream disturbance anomaly. If the anomaly type is protection communication anomaly, mark the anomaly type to be back-replaced as protection communication anomaly. Write the marking results into the anomaly type to be back-replaced list.

[0142] Based on the type of anomaly to be returned, the corresponding location of action is found from the preset mapping table: measurement point anomaly is mapped to the capsule measurement isolation port, topology status anomaly is mapped to the electrical connection edge in the dynamic electrical topology confidence graph, equipment body anomaly is mapped to the capsule conservation ledger port or thermoelectric deviation port, upstream disturbance anomaly is mapped to the boundary node propagation port, and protection communication anomaly is mapped to the capsule protection latching port.

[0143] Write the obtained action location into the list of action locations to be queried, create an index item for the same node-exception type tuple, and record the pointer of the supervised object to be queried, the exception type to be queried, and the action location to be queried.

[0144] After traversing all node-exception type pairs, output the list of supervised objects to be traversed, the list of exception types to be traversed, and the list of positions to be traversed.

[0145] The locations to be replaced include nodes, connecting edges, and ports in the electrical causal residual capsule array in the dynamic electrical topology credibility graph;

[0146] Virtual back-substitution processing is performed according to the type of anomaly to be back-substituted. When the anomaly to be back-substituted is a measurement point anomaly, measurement isolation or substitution input processing is performed on the corresponding measurement port. When the anomaly to be back-substituted is a topology state anomaly, the effective state and edge confidence of the corresponding node or connection edge are corrected. When the anomaly to be back-substituted is a device body anomaly, a body anomaly mark is written to the corresponding electrical causal residual capsule and its conservation ledger port or thermoelectric deviation port is updated. When the anomaly to be back-substituted is an upstream disturbance, a disturbance input is written to the corresponding boundary node. When the anomaly to be back-substituted is a protection communication anomaly, protection signal isolation or communication state correction processing is performed on the corresponding protection latching port.

[0147] After completing the virtual back-substitution process, the electrical causal residual capsule array is reconnected based on the back-substituted dynamic electrical topology credibility graph. The conserved residual, thermoelectric deviation residual, protection logic residual, measurement credibility residual, silent violation residual, and propagation residual are regenerated through the reconnected electrical causal residual capsule array.

[0148] The total amount of conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, propagation residuals, and topological reliability residuals determined by node and edge reliability are calculated separately after back-substitution. Each residual is then weighted and summed according to a preset weight to obtain the total network residual contradiction value after back-substitution. Specifically, the total amounts of conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, propagation residuals, and topological reliability residuals determined by node and edge reliability are calculated separately as follows:

[0149] Traverse the electrical causal residual capsule array, and accumulate the conservation residual values ​​of all monitored objects within the current sampling period to obtain the total conservation residual. Accumulate the thermoelectric deviation residual, protection logic residual, measurement reliable residual, silent violation residual, and propagation residual in the same way to obtain the corresponding total residual. Simultaneously traverse the dynamic electrical topology reliability graph, and record a topology reliable residual event for each node or edge with a node reliability lower than 0.5 or an edge reliability lower than 0.5. Accumulate the number of events to obtain the total topology reliable residual.

[0150] Set fixed weights for each residual total: conservation residual weight 0.20, thermoelectric deviation residual weight 0.15, protection logic residual weight 0.15, measurement reliable residual weight 0.10, silent violation residual weight 0.10, propagation residual weight 0.10, and topological reliable residual weight 0.20.

[0151] Multiply each total residual by the weights mentioned above, and then add the results together to obtain the total network residual discrepancy value after back substitution.

[0152] The difference between the total network residual conflict value before and after virtual back-substitution is calculated. This difference is then compared to the total network residual conflict value before virtual back-substitution to obtain the residual resolution rate for the corresponding explainable resolution candidate. Specifically, the residual resolution rate for the corresponding explainable resolution candidate is as follows:

[0153] Read the network-wide residual contradiction value before virtual back-substitution and record it as the original residual value; then read the network-wide residual contradiction value after virtual back-substitution and record it as the back-substitution residual value.

[0154] Subtract the back-substitution residual from the original residual value to obtain the residual reduction amount;

[0155] Divide the residual reduction amount by the original residual value to obtain the residual reduction percentage in percentage form.

[0156] If the residual reduction amount is positive and the residual reduction ratio is not less than 10%, the ratio is written to the residual reduction rate field of the corresponding interpretable resolution candidate; if the residual reduction amount is not positive or the residual reduction ratio is less than 10%, the residual reduction rate field is written to zero.

[0157] In this embodiment, the step of outputting the anomaly source, anomaly type, anomaly impact range, and corresponding regulatory strategy in the power transmission and transformation system based on the minimum residual resolution set includes:

[0158] Read the residual resolution rate corresponding to each explainable resolution candidate, filter the explainable resolution candidates according to the residual resolution rate threshold, and retain the candidates whose residual resolution rate reaches the residual resolution rate threshold as the candidate resolution options;

[0159] A topology constraint verification is performed on the candidate to be eliminated. If the candidate does not generate new invalid electrical connections, does not destroy the valid connection relationship in the dynamic electrical topology confidence graph, and does not make the edge confidence lower than the preset edge confidence threshold after the candidate is substituted back, the candidate is determined to meet the topology constraint conditions.

[0160] The candidate elimination option is checked for protection logic constraints and silent constraints. When the protection action status after the candidate option is replaced is consistent with the protection closing port of the corresponding supervised object, and no abnormal response is found in the node in the silent state in the dynamic electrical topology confidence diagram, the candidate option is determined to meet the protection logic constraints and silent constraints.

[0161] Candidates that simultaneously satisfy the residual resolution rate threshold, topological constraints, protection logic constraints, and quiescent constraints are combined and screened in order of increasing number of candidates and increasing regulatory cost. The combination of candidates with the fewest number of candidates and the lowest regulatory cost is determined as the minimum residual resolution set.

[0162] The anomaly source and type are determined based on the node-anomaly type tuple in the minimum residual resolution set. The anomaly impact range is determined based on the trusted connection path of the anomaly source in the dynamic electrical topology trust graph. Corresponding regulatory strategies are generated based on the anomaly type. These strategies include measurement point verification strategies, topology review strategies, equipment maintenance strategies, upstream disturbance handling strategies, and protection communication inspection strategies. Specifically, the determination of the anomaly source and type based on the node-anomaly type tuple in the minimum residual resolution set is as follows:

[0163] Traverse each node-anomaly type tuple in the minimum residual resolution set, read the node number in the tuple, use the capsule index table to locate the electrical causal residual capsule to which the node belongs, parse the capsule metadata area to obtain the regulatory object type, primary wiring number and geocode, and determine the node as the anomaly source accordingly.

[0164] Read the anomaly type field in the tuple and use it directly as the anomaly type corresponding to the anomaly source; if the tuple contains multiple anomaly type labels, sort them according to the values ​​of conservation residual, thermoelectric deviation residual, protection logic residual, measurement reliable residual, silent violation residual and propagation residual, and select the anomaly type with the largest residual amplitude as the final anomaly type.

[0165] If multiple anomaly type labels conflict for the same anomaly source, the priority order is as follows: device body anomaly, topology status anomaly, measurement point anomaly, upstream disturbance anomaly, and protection communication anomaly. The anomaly type with the highest priority is retained.

[0166] Write the identified anomaly sources and corresponding anomaly types into the anomaly confirmation table. The fields include the unique ID of the regulated object, the node number, the anomaly type label, the timestamp, and the residual resolution rate.

[0167] Based on the anomaly type, a corresponding regulatory policy is generated, specifically as follows:

[0168] When the anomaly type is a measurement point anomaly, a new work order is created in the measurement point verification queue. The content includes verifying the integrity of the secondary circuit wiring of the current measurement point, comparing the output of the backup current transformer and performing three time synchronization checks, issuing a remote communication command, retrieving the original waveform of the measurement point for 300 consecutive seconds and aligning it with the reference measurement point. If the alignment error is greater than 3%, the measurement point status is set to frozen and the dispatch terminal prompts the use of the backup value.

[0169] When the anomaly type is topology status anomaly, the topology review process is triggered, the remote signaling of the associated switch is set to the verification state, the control center is sent to check the wiring diagram and the on-site label, the UAV inspection subsystem is called to obtain infrared images of the line and switch, the temperature difference between the switch knife edge and the adjacent live conductor is compared, if the temperature difference is less than 2℃, it is marked as actual closing, and the topology model is pushed to the duty officer for a suggestion to update.

[0170] When the anomaly type is equipment body anomaly, an equipment maintenance work order is automatically generated, requiring the station duty officer to perform infrared temperature measurement, insulation resistance test and oil analysis on the equipment on the next maintenance day. In the load management system, the abnormal equipment is set to an 80% load limit operation mode, and a backup power supply switching plan is generated to ensure load transfer.

[0171] When the anomaly type is upstream disturbance anomaly, the system will link with the monitoring platform of the adjacent dispatch area to send a query command for the voltage disturbance of the upper bus and the power angle of the parallel line. The affected lines and equipment will be highlighted on the dispatch graphical interface, and suggestions will be made to postpone reclosing and high load start-up and shutdown operations. If the voltage stability of the upper bus recovers to within ±3% of the rated value, the system will automatically cancel the suggestion.

[0172] When the anomaly type is protection communication anomaly, a channel loopback self-test command is sent to the relay protection backend, and the delay of the main and backup channels is compared. If the delay of the main channel is higher than 20ms and the delay of the backup channel is normal, the protection device is instructed to exit the main channel and enter the backup channel. A maintenance work order is pushed, requiring the communication professionals to troubleshoot the optical transceiver, power supply and SDH ring network.

[0173] Example 1: During a continuous power transmission and transformation monitoring cycle, the system receives operational data from a 220kV power transmission and transformation system. The monitored objects include 2 busbar sections, 2 main transformers, 4 lines, 18 circuit breakers, 24 disconnect switches, 96 measuring points, and 12 protection devices. One sampling window contains 3600 sampling points, with a sampling interval of 1 second. The raw data includes approximately 2.86 million records of voltage, current, active power, reactive power, voltage phase angle, switch remote signaling, protection actions, equipment temperature, measuring point communication status, and primary wiring data. After data access, it was found that the current measuring point missing rate was 0.42%, there were 17 records of switch remote signaling jitter, 23 abnormal spikes in temperature measuring points, and 312 records with communication delays exceeding 100ms.

[0174] The system first preprocesses the multi-source operating data. All data is standardized to a 1-second sampling period. Missing time points are filled by averaging the three adjacent sampling points. Data exceeding 1.2 times the equipment's rated range are marked as abnormal and removed. Before processing, the maximum fluctuation of the line current data was ±96A; after processing, it stabilized at ±18A. The maximum deviation of the measurement point timestamps decreased from 236ms to 18ms. After normalization, voltage, current, power, temperature, and phase angle data are all converted to the 0–1 interval, forming a standard dataset.

[0175] The system constructs a dynamic electrical topology diagram based on a standard dataset. It reads the busbar, line, transformer, circuit breaker, and disconnector numbers from primary wiring data, generating 50 equipment nodes; 96 measurement point nodes from the measurement point binding table; and 12 protection nodes from the protection zone table. Based on equipment connection relationships, it generates 128 electrical connection edges, 96 measurement-related edges, and 42 protection-related edges. After scanning the switch status, 17 of the 18 circuit breakers are closed, and 1 is open; 22 of the 24 disconnectors are closed, and 2 are open. The system marks the 7 electrical connection edges controlled by the open switch as invalid, retaining 121 valid electrical connection edges.

[0176] In the improved node2vec model processing stage, the system extracts the power flow direction vector, voltage phase angle difference, and switch status reliability for each valid electrical connection edge with a preset refresh cycle of 60 seconds. Taking the first-end node to the last-end node of a certain line as an example, the real-time power flow direction is consistent with the specified power flow direction, and the power flow direction consistency weight is 1.000; the voltage phase angle difference between the two ends is 3.4°, and the phase angle consistency weight is 1.000; the corresponding circuit breaker remote signaling is complete and has not jittered within 60 seconds, and the switch reliability weight is 1.000. The switch remote signaling of another bus tie edge jittered 3 times within 60 seconds, and the switch reliability weight was set to 0.300. After fusing the structural transfer weight, power flow direction consistency weight, phase angle consistency weight, and switch reliability weight, the system generates the improved node2vec random walk transfer probability function. 80 random walks are performed on each node, with each walk having a length of 40, ultimately generating a 64-dimensional node embedding vector for 158 nodes. After comparing with the historical normal embedding vector, the confidence level of the main transformer node is 0.94, the confidence levels of the two bus segments are 0.96 and 0.95 respectively, the confidence level of the current measuring point at the beginning of a certain line drops to 0.57, and the confidence level of the measurement association edge between this measuring point and the line node is 0.54. The dynamic electrical topology confidence graph is thus generated.

[0177] During the construction phase of the electrical causality residual capsule array, the system creates an electrical causality residual capsule for each monitored object. Line capsules include a starting voltage port, a starting current port, an ending voltage port, an ending current port, a power port, a temperature port, a protection engagement port, and a residual propagation port; transformer capsules include a high-voltage side port, a medium-voltage side port, a low-voltage side port, a tap changer port, an oil temperature port, and a differential protection port. The system connects capsules according to node confidence and edge confidence. Measurement-related edges with an edge confidence lower than 0.60 are marked as low-confidence connections but are still retained for anomaly analysis. The final result is an electrical causality residual capsule array containing 158 capsules, 121 electrical propagation channels, 96 measurement-related channels, and 42 protection-related channels.

[0178] During this regulatory period, after relative time T0, the current at the beginning of a certain line suddenly increased from 506A to 692A, lasting for 14 seconds; the current at the end only changed from 502A to 509A; the active power at the beginning of the line suddenly increased from 165MW to 224MW, but the active power at the end remained between 163MW and 167MW; the temperature at the line joint increased from 46.3℃ to 46.8℃, without a temperature rise matching the 692A current; neither distance protection nor overcurrent protection was activated. The capsule array calculation showed that the line's conserved residual increased from 2.9MW to 22.4MW, the thermoelectric deviation residual was 4.6℃, the protection logic residual count was 4, the measurement reliability residual count was 9, the silent violation residual was 0, and the propagation residual count was 5. The system arranged the above results according to the regulated object, residual type, and sampling time, generating multi-type residual tensors with a tensor size of 158×6×3600.

[0179] The system couples various types of residual tensors with node and edge credibility, performing local density clustering on the dynamic electrical topology credibility graph. A current measurement point at the beginning of a line, a line capsule, and the power port of the beginning bus are grouped into the same anomalous cluster. This cluster contains 5 monitored objects with a local density of 3.72, exceeding the first density threshold of 2.50. The system performs residual minimum cut entropy reduction processing on the anomalous cluster. When the measurement association edge between the current measurement point at the beginning and the line capsule is used as the target cutting edge, the residual distribution entropy decreases from 2.84 to 1.31, an entropy reduction of 1.53. When the line itself is used as the cutting object, the entropy reduction is only 0.48. Therefore, the system generates three interpretable resolution candidates: current measurement point at the beginning—measurement point anomaly, line itself—equipment itself anomaly, and measurement association edge at the beginning—measurement point mapping anomaly.

[0180] During the virtual back-substitution phase, the system first performs measurement isolation for anomalies in the head-end current measuring point, and estimates the replacement current value of 508A using the end current, bus power balance, and historical load curves. After back-substitution, the conservation residual decreased from 22.4MW to 3.2MW, the protection logic residual decreased from 4 to 0, the measurement reliability residual decreased from 9 to 1, the propagation residual decreased from 5 to 1, and the network-wide residual inconsistency value decreased from 100.0 to 27.6, with a residual resolution rate of 72.4%. When back-substitution for anomalies in the line body and equipment body, the system writes additional loss markers into the line capsule. After back-substitution, the network-wide residual inconsistency value is 75.8, with a residual resolution rate of 24.2%. When back-substitution for anomalies in the head-end measurement associated edge and measuring point mapping, the system attempts to rebind the measuring points, but introduces new silent violation residuals. The network-wide residual inconsistency value is 63.5, with a residual resolution rate of 36.5%.

[0181] The system determines the minimum residual resolution set based on residual resolution rate, topology constraints, protection logic constraints, and quiescent constraints. Since the residual resolution rate is highest for the current measurement point at the beginning of the line, and the anomaly does not disrupt the topology connection, violate the protection logic, or generate a quiescent violation, the system ultimately outputs the anomaly source as the current measurement point at the beginning of the line, the anomaly type as measurement point anomaly, the anomaly impact range as the measurement link at the beginning of the line and its associated alarms, and the monitoring strategy is to check the binding relationship between the current transformer secondary circuit, the acquisition device, and the measurement point, reducing the weight of the measurement point in the monitoring to 0.30.

[0182] To verify the effectiveness, the system used 12,000 training samples, 3,000 validation samples, and 3,000 test samples, including 3,200 normal samples, 1,800 samples with measurement point anomalies, 1,300 samples with switch remote signaling anomalies, 1,200 samples with measurement point mapping anomalies, 2,100 samples with equipment body anomalies, 1,400 samples with upstream disturbances, and 1,000 samples with protection communication anomalies. Traditional methods, using static topology and threshold alarms, achieved an anomaly source identification accuracy of 81.4%, an anomaly type identification accuracy of 78.9%, a false alarm rate of 18.6% for equipment body faults, and an accuracy of 73.1% for measurement point anomaly identification, with an average anomaly confirmation time of 42.5 seconds. The corresponding results for the method of this invention are 94.7%, 93.5%, 5.2%, 95.8%, and 18.7 seconds, respectively. In 300 mixed samples of abnormal current at the first end, the traditional method correctly identified 190 groups, while the present invention correctly identified 278 groups. Among the 120 groups of abnormal actual measurement points, 112 groups were correctly identified, and among the 90 groups of abnormal actual equipment, 85 groups were correctly identified.

[0183] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A power transmission and distribution system monitoring method based on artificial intelligence, characterized by, include: Acquire multi-source operation data from the power transmission and transformation system, preprocess the multi-source operation data, and obtain a standard dataset; A dynamic electrical topology graph is constructed based on a standard dataset. The power flow direction vector, voltage phase angle difference, and switch state confidence extracted from the standard dataset within a preset refresh cycle are used as transfer weight factors to modify the random walk transfer probability function in the improved node2vec model. The node embedding update is performed on the dynamic electrical topology graph through the improved node2vec model, and the node confidence and edge confidence are calculated to obtain a reliable dynamic electrical topology graph. Based on the dynamic electrical topology credibility graph, each regulated object is mapped as an electrical causal residual capsule. Each electrical causal residual capsule is connected according to the node credibility and edge credibility to form an electrical causal residual capsule array, generating multiple types of residual tensors. Couple multiple types of residual tensors, node credibility and edge credibility, perform embedded spatial local density clustering on dynamic electrical topology credibility graph to identify anomalous clusters, and use residual minimum cut entropy reduction strategy to generate node-anomaly type binary set for each anomalous cluster to obtain interpretable resolution candidates. The explainable and resolvable candidate options are virtually substituted back into the dynamic electrical topology credibility graph and the electrical causal residual capsule array, and the residual contradiction value of the whole network after the substitution is recalculated to generate the residual resolution rate of the corresponding candidate options; The minimum residual resolution set is determined based on the residual resolution rate, topological constraints, protection logic constraints, and quiescent constraints. Based on the minimum residual resolution set, the abnormal source, abnormal type, abnormal impact range, and corresponding regulatory strategy in the power transmission and transformation system are output. 2.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The multi-source operating data includes voltage data, current data, active power data, reactive power data, voltage phase angle data, frequency data, switch status data, protection action data, equipment temperature data, partial discharge data, gas pressure data, surge arrester leakage current data, circuit breaker mechanical characteristic data, measuring point number data, measuring point binding relationship data, communication status data, ambient temperature data, humidity data, wind speed data, primary wiring data, equipment connection relationship data, and protection zone relationship data. 3.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The preprocessing of multi-source operating data to obtain a standard dataset includes: uniformly timestamping the multi-source operating data; aligning the time according to a preset sampling period; removing outliers that exceed the rated operating range of the equipment; completing missing data; unifying the mapping relationship between measurement point numbers and equipment numbers; converting data from different sources into a unified data format; and normalizing voltage data, current data, active power data, reactive power data, voltage phase angle data, equipment status data, and environmental data to obtain the standard dataset. 4.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The obtained dynamic electrical topology reliability map includes: Based on the primary wiring data, equipment connection relationships, switch status data, measurement point binding relationships, and protection zone relationships in the standard dataset, a dynamic electrical topology diagram containing equipment nodes, measurement point nodes, protection nodes, and electrical connection edges is constructed, and the effective status of each electrical connection edge is determined according to the switch opening and closing status. Within a preset refresh cycle, the power flow direction vector, voltage phase angle difference, and switch status confidence level corresponding to each valid electrical connection edge are extracted from the standard dataset. The power flow direction vector is converted into power flow direction consistency weight, the voltage phase angle difference is converted into phase angle consistency weight, and the switch status confidence level is converted into switch confidence weight. For a random walk process of the improved node2vec model, the current walking node, the previous walking node, and the candidate transition nodes of the current walking node are determined. The candidate transition nodes are the adjacent nodes that are connected to the current walking node through an effective electrical connection edge. Based on the return and input / output parameters of the node2vec model, the structural transfer weight of each candidate transfer node relative to the previous roaming node is determined. The structural transfer weight, power flow direction consistency weight, phase angle consistency weight, and switch confidence weight are then fused to obtain the comprehensive transfer weight of each candidate transfer node. The combined transfer weights of all candidate transfer nodes corresponding to the same current walk node are summed, and the ratio of the combined transfer weight of each candidate transfer node to the summation result is determined as the random walk transfer probability of the candidate transfer node. The random walk transfer probabilities of each candidate transfer node constitute the random walk transfer probability function of the improved node2vec model. The node walking sequence of the dynamic electrical topology graph is generated based on the random walk transition probability function. The node embedding sequence is updated by improving the node2vec model to obtain the node embedding vector of each node. The node credibility and edge credibility are calculated based on the node embedding vector, historical normal embedding vector, effective state of electrical connection edge and time-varying edge weight to obtain the dynamic electrical topology credibility graph. 5.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The formation of the electrical causal residual capsule array generates multiple types of residual tensors, including: Based on the dynamic electrical topology credibility graph, the regulatory objects in the power transmission and transformation system are determined. The busbar, line, transformer, circuit breaker, disconnector, cable joint, bushing, surge arrester, protection device and measuring point are treated as independent regulatory objects. A corresponding electrical causal residual capsule is created for each regulatory object. The port configuration of the corresponding electrical causal residual capsule is determined according to the equipment type of the regulated object. The electrical causal residual capsule corresponding to the busbar is configured as an incoming port, an outgoing port, a bus tie port, a voltage phase angle port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the line is configured as a head electrical port, an end electrical port, a temperature port, a measurement port, and a protection port. The electrical causal residual capsule corresponding to the transformer is configured as a high-voltage side port, a medium-voltage side port, a low-voltage side port, a tap changer port, a temperature port, a cooling status port, a measurement port, and a protection port. The voltage, current, active power, reactive power, voltage phase angle, equipment temperature, protection action, measurement point, communication status, and environmental data that are consistent with the regulated object identifier in the standard dataset are connected to the port of the corresponding electrical causal residual capsule. The node credibility and edge credibility in the dynamic electrical topology credibility graph are written into the corresponding electrical causal residual capsule. According to the effective connection relationship in the dynamic electrical topology confidence graph, each electrical causal residual capsule is connected. When there is an effective electrical connection edge between two monitored objects and the confidence of the corresponding edge meets the preset connection conditions, a residual propagation connection is established between the two electrical causal residual capsules. When there is a binding relationship between the monitored object and the measuring point or protection device, a measurement association connection or protection association connection is established between the corresponding electrical causal residual capsules to form an electrical causal residual capsule array. By generating conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, and propagation residuals for each regulated object through an electrical causal residual capsule array, and arranging them according to the regulated object, residual type, and sampling time to generate multi-type residual tensors. 6.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The obtained interpretable resolution candidates include: We obtain multiple types of residual tensors, node credibility, edge credibility, and node embedding results obtained by the improved node2vec model during the node embedding update process. We then concatenate the residual type, residual magnitude, node credibility, associated edge credibility, and node embedding results corresponding to the same regulatory object to form the clustering features of the regulatory object. In the dynamic electrical topology credibility graph, the cluster feature distance of the regulated objects is used as the embedding space distance, and the edge credibility is used as the topological connection constraint to determine the effective neighborhood of each regulated object. The local density of the regulated objects is determined based on the number of residual similar regulated objects in the effective neighborhood. The regulatory objects whose local density reaches the first density threshold and whose multi-type residual tensors reach the first residual threshold are identified as cluster cores. The regulatory objects that are within the embedding space distance of the cluster cores and are connected by trusted edges are merged into the same cluster to obtain the abnormal cluster. For each abnormal cluster, candidate splitting edges are determined based on the residual differences between adjacent monitored objects within the cluster, edge credibility, and residual propagation direction. The distribution changes of conserved residuals, thermoelectric deviation residuals, protection logic residuals, measurement credibility residuals, silent violation residuals, and propagation residuals in each sub-cluster before and after splitting are statistically analyzed. The candidate splitting edge whose residual distribution entropy decreases to the first entropy reduction threshold and has the lowest splitting cost is determined as the target splitting edge. The abnormal clusters are segmented according to the target segmentation edge. The regulatory objects with the highest residual concentration after segmentation and whose node credibility and edge credibility meet the abnormal location conditions are identified as abnormal candidate nodes. The abnormal type is determined according to the dominant residual type corresponding to the abnormal candidate node. The abnormal candidate node and the abnormal type are combined to generate a set of node-abnormal type binary pairs to obtain the interpretable resolution candidate. 7.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The residual resolution rate for generating the corresponding candidate options includes: Read the set of node-anomaly type pairs in the interpretable resolution candidate, and determine the regulatory object to be back-submitted, the anomaly type to be back-submitted, and the location of the back-submitted action based on each node-anomaly type pair. Virtual back-substitution processing is performed according to the type of anomaly to be back-substituted. When the anomaly to be back-substituted is a measurement point anomaly, measurement isolation or substitution input processing is performed on the corresponding measurement port. When the anomaly to be back-substituted is a topology state anomaly, the effective state and edge confidence of the corresponding node or connection edge are corrected. When the anomaly to be back-substituted is a device body anomaly, a body anomaly mark is written to the corresponding electrical causal residual capsule and its conservation ledger port or thermoelectric deviation port is updated. When the anomaly to be back-substituted is an upstream disturbance, a disturbance input is written to the corresponding boundary node. When the anomaly to be back-substituted is a protection communication anomaly, protection signal isolation or communication state correction processing is performed on the corresponding protection latching port. After completing the virtual back-substitution process, the electrical causal residual capsule array is reconnected based on the back-substituted dynamic electrical topology credibility graph. The conserved residual, thermoelectric deviation residual, protection logic residual, measurement credibility residual, silent violation residual, and propagation residual are regenerated through the reconnected electrical causal residual capsule array. The total amount of conservation residuals, thermoelectric deviation residuals, protection logic residuals, measurement reliability residuals, silent violation residuals, propagation residuals, and topological reliability residuals determined by node reliability and edge reliability are calculated separately after back-substitution. The total amount of each residual is weighted and summed according to the preset weights to obtain the total residual contradiction value of the whole network after back-substitution. The difference between the total network residual contradiction value before virtual back-substitution and the total network residual contradiction value after back-substitution is calculated. The ratio of the difference to the total network residual contradiction value before virtual back-substitution is calculated to obtain the residual resolution rate of the corresponding interpretable resolution candidate. 8.The power transmission and distribution system monitoring method based on artificial intelligence according to claim 1, wherein, The method of outputting the anomaly source, anomaly type, anomaly impact range, and corresponding regulatory strategy in the power transmission and transformation system based on the minimum residual resolution set includes: Read the residual resolution rate corresponding to each explainable resolution candidate, filter the explainable resolution candidates according to the residual resolution rate threshold, and retain the candidates whose residual resolution rate reaches the residual resolution rate threshold as the candidate resolution options; A topology constraint verification is performed on the candidate to be eliminated. If the candidate does not generate new invalid electrical connections, does not destroy the valid connection relationship in the dynamic electrical topology confidence graph, and does not make the edge confidence lower than the preset edge confidence threshold after the candidate is substituted back, the candidate is determined to meet the topology constraint conditions. The candidate elimination option is checked for protection logic constraints and silent constraints. When the protection action status after the candidate option is replaced is consistent with the protection closing port of the corresponding supervised object, and no abnormal response is found in the node in the silent state in the dynamic electrical topology confidence diagram, the candidate option is determined to meet the protection logic constraints and silent constraints. Candidates that simultaneously satisfy the residual resolution rate threshold, topological constraints, protection logic constraints, and quiescent constraints are combined and screened in order of increasing number of candidates and increasing regulatory cost. The combination of candidates with the fewest number of candidates and the lowest regulatory cost is determined as the minimum residual resolution set. The abnormal source and the abnormal type are determined according to the node-exception type binary tuple in the minimum residual resolution set, the abnormal influence range is determined according to the trusted connection path of the abnormal source in the dynamic electrical topology trusted graph, and the corresponding monitoring strategy is generated according to the abnormal type.

Images