An intelligent substation operation state evaluation system based on artificial intelligence

Abstract

The application relates to the technical field of power operation management, and discloses an intelligent substation operation state evaluation system based on artificial intelligence. The system is deployed in a power grid dispatching automation master station and a substation field monitoring network cooperative environment, unified structured access and semantic standardization processing are performed on multi-source monitoring data on the master station side and the station end side, a secondary equipment object model and an association graph atlas covering equipment entities, communication links and time-dependent relationships are constructed, a self-adaptive dynamic graph structure model with space-time coupling capability is established on the basis, dynamic adjacency weight mechanism, cross-time window state propagation mechanism and internal cycle self-supervision updating strategy are used, dynamic intelligent judgment and abnormal propagation correlation reasoning of secondary equipment operation states are realized, and therefore the accuracy, stability and interpretability of substation operation state analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of power operation management technology, and in particular to an intelligent evaluation system for substation operation status based on artificial intelligence. Background Technology

[0002] With the advancement of new power system construction and the continuous improvement of substation digitalization and networking, a large number of secondary functional equipment such as relay protection devices, measurement and control devices, remote terminals, communication management units and time synchronization devices are centrally monitored and data aggregated through dispatch automation systems, forming multi-source high-frequency time-series data including telemetry, remote signaling, remote control, SOE events and communication messages. Existing intelligent technologies have attempted to use machine learning or deep learning methods to classify, predict, or identify anomalies in single-device operational data, which has improved the automation level of equipment status assessment to some extent. However, most solutions use independent devices or static feature vectors as modeling objects, focusing on the analysis of single time-slice data. They lack a unified graph-based model of structural dependencies, communication topology relationships, and timing link relationships between secondary devices, and fail to dynamically depict the state propagation mechanism across devices, intervals, and links during time evolution. At the same time, existing methods are often trained under fixed topology or static adjacency relationships, making it difficult to adapt to the dynamic characteristics of relationships changing over time during operation. They have limited robustness to remote signal jitter, short-term link fluctuations, and multi-alarm coupling disturbances, and their generalization ability is insufficient in weakly labeled scenarios or scenarios with scarce abnormal samples. In addition, multi-source monitoring data in the collaborative environment between the master station and the station end have problems such as time alignment error, inconsistent identifier mapping and semantic expression. Existing intelligent models are usually trained directly based on preprocessed feature inputs, lacking a unified and systematic design from data governance, object modeling to graph structure construction, resulting in insufficient support for cross-device correlation interpretation and anomaly propagation tracing of model output results. Summary of the Invention

[0003] This invention addresses the challenges of heterogeneous and dispersed multi-source monitoring data, complex and dynamically evolving structural dependencies between equipment, and the difficulty in uniformly modeling the cross-device and cross-link propagation characteristics of abnormal states during the operation of substation secondary systems. It constructs an AI-based intelligent assessment system for substation operation status. In a collaborative environment between the dispatch automation master station and the substation field monitoring network, this system performs unified structured modeling of multi-source monitoring data, constructing a secondary equipment object model and relational graph covering equipment entities, communication links, timing dependencies, and configuration constraints. Based on this, an adaptive dynamic graph structure model with spatiotemporal coupling capabilities is established. Dynamic adjacency weights and cross-time window state propagation mechanisms are used to dynamically express the evolution of the operation status. Furthermore, the system combines internal loop self-supervised updates and graph structure reconstruction constraints to enhance the model's ability to characterize and generalize complex abnormal patterns. This enables multi-granular intelligent judgment of the secondary equipment operation status and abnormal propagation correlation reasoning, improving the intelligent analysis and risk identification capabilities of substation operation management.

[0004] This invention provides an intelligent substation operation status assessment system based on artificial intelligence, deployed in a collaborative environment of power grid dispatch automation master station and substation field monitoring network. It is used to intelligently analyze the operation status of secondary equipment in substation based on monitoring information from the dispatch automation system. The system includes: a data access module, a data governance module, an equipment modeling module, a status feature extraction module, a model training and management module, an operation status assessment module, and an anomaly correlation reasoning module.

[0005] The data access module accesses monitoring information from the dispatch automation system and forms a unified structured monitoring data stream. It also performs protocol parsing, device identifier mapping, tag standardization, and unified data encapsulation processing on data from different sources to obtain standardized raw monitoring data.

[0006] The data governance module performs timestamp correction, outlier identification, missing data completion, duplicate event resolution, status jitter suppression, and multi-source time series alignment on standardized monitoring raw data to generate standardized monitoring time series data for secondary equipment status analysis.

[0007] The equipment modeling module constructs object models of secondary equipment in substations and their relationship graphs.

[0008] The status feature extraction module extracts event time sequence segment features, operational health features, command response features, alarm co-occurrence features, communication link quality features, configuration consistency features, and cross-device linkage features from standardized monitoring time sequence data to form a secondary device status feature vector;

[0009] The model training and management module constructs an adaptive dynamic graph structure model. Based on historical standardized monitoring time-series data, historical secondary equipment state feature vectors, and corresponding historical operating state labels, the adaptive dynamic graph structure model is trained offline to generate a trained adaptive dynamic graph structure model. The adaptive dynamic graph structure model includes a dynamic graph neural network for graph structure modeling, a long short-term memory network for time-series modeling, a training inner loop unit during testing, and a variational graph autoencoder network for graph structure reconstruction constraints.

[0010] The operation status assessment module calls the trained adaptive dynamic graph structure model to perform online reasoning processing on the current secondary equipment status feature vector, and outputs the secondary equipment operation status judgment result. The secondary equipment operation status judgment result includes normal status, abnormal fluctuation status, communication abnormal status, alarm abnormal status, action mismatch status, clock abnormal status, interlocking abnormal status and suspected fault status.

[0011] The anomaly correlation reasoning module combines the secondary equipment object model, correlation graph, and operation status judgment results to perform correlation reasoning on the anomaly propagation relationships across devices, bays, and communication links, generating anomaly correlation reasoning results. Based on the operation status judgment results and anomaly correlation reasoning results, it generates equipment-level, bay-level, and station-level status assessment results, and outputs alarm levels, anomaly types, evidence event sets, suggested verification paths, and handling suggestions to operation and management personnel to support substation operation management and risk control decisions.

[0012] Furthermore, the state feature extraction module includes the following feature construction components:

[0013] Based on the remote signaling change records, SOE sequence event records and protection action information contained in the standardized monitoring time series data, related events in the same interval are aggregated according to the preset time window to construct an event sequence window; within the event sequence window, the event occurrence order features, adjacent event interval time features, unit time action trigger frequency features, and abnormal event combination pattern features are extracted to generate event time series segment features corresponding to the device.

[0014] Based on telemetry data and device self-test information in standardized monitoring time series data, operational statistical features are constructed within a preset time window. Features of electrical quantity fluctuation amplitude, periodic change, over-limit ratio, self-test anomaly count, and continuous stable running time are extracted to generate operational health features that characterize the stability and health status of equipment operation.

[0015] Based on the remote control issuance records and corresponding status feedback records in the standardized monitoring time series data, a command-response matching relationship is established under a unified time axis. Command execution statistical features are constructed, and command response success rate features, response delay distribution features, non-response count features, and abnormal feedback type features are extracted to generate command response features, which are used to characterize the consistency, timeliness, and abnormal response behavior during the execution of equipment control commands.

[0016] Based on alarm information in standardized monitoring time series data, an alarm statistical model is constructed within a preset time window. Alarm density features, alarm count features per unit time, alarm duration features, and alarm co-occurrence matrix features are extracted to generate alarm co-occurrence features, which are used to characterize the concentration, persistence, and multi-alarm co-occurrence patterns of equipment alarms.

[0017] Based on the communication link status information and network status records in the standardized monitoring time series data, link operation statistical analysis is performed in the time dimension to extract link interruption frequency characteristics, link recovery time characteristics, packet loss rate characteristics, port anomaly count characteristics, and network jitter characteristics, and generate communication link quality characteristics to characterize the stability and reliability level of the device's communication channel.

[0018] Based on the pressure plate status information, setpoint area information and clock synchronization status information in the standardized monitoring time sequence data, and combined with the configuration constraint rules in the equipment object model, configuration status consistency features are constructed, and function activation status identifier features, setpoint area switching record features, pressure plate activation / deactivation consistency features and clock synchronization stability features are extracted to generate configuration consistency features, which are used to characterize whether the current operating configuration of the equipment meets the engineering logic constraints and protection function activation conditions.

[0019] Based on the relationship graph and standardized monitoring time series data, we construct the inter-device collaboration features in the topology and time dimensions, extract the matching features of protection actions and measurement and control feedback, the time difference features of event propagation between devices, the time synchronization dependency consistency features, and the link anomaly linkage features, and generate cross-device linkage features to characterize the collaborative operation behavior and anomaly propagation patterns between different devices.

[0020] Furthermore, the offline training process of the model training and management module for the adaptive dynamic graph structure model includes: constructing a training sample set containing historical standardized monitoring time-series data, historical secondary equipment state feature vectors, and historical operating state labels; generating historical spatiotemporal dynamic graph representation sequences corresponding to the training samples based on the training sample set; performing graph augmentation processing on the historical spatiotemporal dynamic graph representation sequences to construct a contrastive loss term, inputting the historical spatiotemporal dynamic graph representation sequences into a variational graph autoencoder network to construct a VGAE loss term, and constructing a classification loss term based on the historical operating state labels; weightedly fusing the contrastive loss term, VGAE loss term, and classification loss term to construct a total loss function; and performing end-to-end backpropagation and joint optimization training based on the total loss function to obtain the trained adaptive dynamic graph structure model.

[0021] Furthermore, the process of calling the trained adaptive dynamic graph structure model to process the secondary equipment state feature vector and output the secondary equipment operating status determination result specifically includes the following steps:

[0022] Step S1: Collect, align and concatenate the current secondary equipment state feature vectors according to a unified time window to obtain the secondary equipment state analysis input tensor, and divide the secondary equipment state analysis input tensor into a time window sequence according to the time dimension; construct a window using the first time window in the time window sequence as the initial graph, and generate the initial node state representation based on the feature sub-vectors of each feature node in the first time window;

[0023] Step S2: For the current time window in the time window sequence, if the current time window is the first time window, the initial node state representation is used as the historical state representation of the current time window; if the current time window is a subsequent time window after the first time window, the node state representation output by the previous time window is used as the historical state representation of the current time window; based on the historical state representation of the current time window, a time propagation edge is established between adjacent time windows of the same feature node, and a dynamic adjacency matrix corresponding to the current time window is constructed; the dynamic adjacency matrix is ​​used to perform message passing and node embedding aggregation operations.

[0024] Step S3: Call the inner loop unit trained during testing on the time window sequence to perform dynamic feature fusion update based on historical dependencies on the secondary device status analysis input tensor, generating the feature tensor updated by the inner loop; input the feature tensor updated by the inner loop into the long short-term memory network in parallel to generate the temporal embedding matrix;

[0025] Step S4: For the dynamic adjacency matrix corresponding to the current time window, construct an interpretability weight matrix with the same dimension as the dynamic adjacency matrix;

[0026] Step S5: In the dynamic graph neural network, for each time window, the dynamic adjacency matrix, interpretable weight matrix and temporal embedding matrix are fused to obtain the spatiotemporal graph representation of that time window. After fusion of time windows piece by piece, a spatiotemporal dynamic graph representation sequence corresponding to the secondary device state feature vector is formed.

[0027] Step S6: Input the spatiotemporal dynamic graph representation sequence into the classification output layer of the trained adaptive dynamic graph structure model to obtain the state probability distribution of each preset operating state category, and output the secondary equipment operating state determination result according to the state probability distribution.

[0028] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0029] This invention achieves structured integration and semantically unified expression of multi-source monitoring data from the main station and station ends by constructing a unified secondary equipment object model and relationship graph. It solves the problems of scattered monitoring information sources, inconsistent equipment identification, and difficulty in uniformly describing relationships. It enables integrated modeling of equipment entities, communication links, timing dependencies, and configuration constraints in the same graph structure, improving the overall organizeability and analyzability of substation secondary system operation data. This provides a complete and accurate data foundation for subsequent operation status assessment and anomaly reasoning, and enhances the system's applicability in complex substation structural environments.

[0030] This invention introduces an adaptive dynamic graph structure model with spatiotemporal coupling representation capabilities, enabling the modeling of the dynamic evolution of structural dependencies between devices over time. This enhances the ability to characterize anomaly propagation patterns across devices, intervals, and links, solving the problem that traditional static modeling cannot reflect the continuous evolution of operating states. By combining a training loop update mechanism during testing with a cumulative state expression structure, the model enhances its ability to identify latent anomalies and complex linkage patterns while maintaining the integrity of the original features, thereby improving the accuracy and stability of operating state determination.

[0031] This invention achieves joint optimization of operational status classification capability and graph structure generation consistency by combining graph comparison enhancement and variational graph autoencoding reconstruction constraints. This enhances the model's generalization ability and robustness in scenarios with weak labels and scarce abnormal samples, enabling the system to output multi-granularity evaluation results at the equipment, bay, and station levels. Furthermore, it generates interpretable anomaly propagation paths and handling suggestions by combining correlation graphs, solving the problem of insufficient support for anomaly tracing and improving the intelligent analysis level and risk control decision support capability of substation operation management. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of a module of an intelligent substation operation status assessment system based on artificial intelligence proposed in this invention.

[0033] Figure 2 This is a convergence curve of each sub-loss during the training process proposed in Example 3. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] Example 1, according to Figure 1 This invention provides an intelligent substation operation status assessment system based on artificial intelligence, deployed in a collaborative environment of power grid dispatch automation master station and substation field monitoring network. It is used to intelligently analyze the operation status of secondary equipment in substation based on monitoring information from the dispatch automation system. The system includes: a data access module, a data governance module, an equipment modeling module, a status feature extraction module, a model training and management module, an operation status assessment module, and an anomaly correlation reasoning module.

[0036] The data access module accesses monitoring information from the dispatch automation system and forms a unified structured monitoring data stream. It performs protocol parsing, device identifier mapping, tag standardization, and unified data encapsulation on data from different sources to obtain standardized raw monitoring data. Specifically, it includes: a dispatch master station data access unit, a station-side message acquisition unit, a protocol parsing unit, a tag standardization unit, and a device mapping unit. The dispatch master station data access unit acquires raw monitoring records containing timestamps, station names, interval names, signal identifiers, signal values, and event types from the EMS / SCADA system, the dispatch automation historical database, and the real-time alarm stream interface, forming the master station-side raw monitoring data stream. The station-side message acquisition unit collects communication messages from secondary equipment within the station, device operating status information, switch port status information, and timing status information through log subscription, forming the station-side raw message data stream. The protocol parsing unit performs protocol identification and field decomposition on the master station-side raw monitoring data stream and the station-side raw message data stream, parsing to obtain structured data records. The data records include: a unified timestamp field, a site identifier field, a device identifier field, a signal type field, a signal value field, an event category field, a source identifier field, and an original message index field; the tag standardization unit is used to map the point number, signal name, address code, and device name in data from different sources according to a unified naming rule to generate standard tag fields. The standard tag fields adopt a hierarchical structure of "site name-voltage level-interval-device-signal item" and establish a correspondence table between the original identifier and the standard tag; the device mapping unit is used to establish a mapping relationship between the standard tag and the on-site secondary equipment entity, logical node, communication port, interval number, and loop number based on the pre-built secondary equipment asset ledger and logical node configuration table, and to supplement each structured data record with a unique device identifier, logical node code, and physical port identifier information; the data encapsulation and output unit is used to encapsulate the structured data records after protocol parsing, tag standardization, and device mapping processing in a unified data format to form standardized monitoring raw data;

[0037] The data governance module performs timestamp correction, outlier identification, missing data completion, duplicate event resolution, status jitter suppression, and multi-source time series alignment on standardized monitoring raw data to generate standardized monitoring time series data for secondary equipment status analysis.

[0038] The equipment modeling module constructs a model of the secondary equipment objects in the substation and its relational graph. The secondary equipment object model uses the substation's secondary equipment as the core modeling object, including: relay protection devices, measurement and control devices, fault recording devices, automatic devices, communication management units, remote terminals, switches, merging units, intelligent terminals, and time synchronization devices. Each equipment entity is configured with a unique equipment identifier, its voltage level, its bay number, functional category identifier, communication port information, logical node identifier, and operating attribute parameters. The relational graph is represented in graph structure form, with nodes including equipment nodes, bay nodes, primary equipment nodes, communication nodes, and event nodes. Edges represent the structural relationships between nodes. These relationships include: 1. Monitoring mapping relationship: the monitoring relationship between secondary equipment and primary equipment, used to represent the monitoring and control relationship of protection or measurement and control devices on the corresponding primary equipment; 2. Collaborative logical relationship: the collaborative relationship between relay protection devices and measurement and control devices, used to represent the logical association between action criteria and remote signaling feedback; 3. Network connection... 4. Connection Relationship: The connection relationship between secondary functional equipment entities and communication network equipment, used to represent the port mapping relationship, link status association relationship, and communication path dependency relationship between relay protection devices, measurement and control devices, fault recording devices, automatic devices, remote terminals, merging units, intelligent terminals, and time synchronization devices and communication management units; 5. Timing Dependency Relationship: The timing dependency relationship between secondary functional equipment entities and clock synchronization sources, used to represent the time synchronization dependency path and timing link level of relay protection devices, measurement and control devices, fault recording devices, automatic devices, remote terminals, merging units, intelligent terminals, and switches to time synchronization devices; 6. Configuration Constraint Relationship: The configuration association relationship between secondary functional equipment entities and their setting area status and pressure plate status, used to represent the function activation status, activation / deactivation status, and setting value switching status of relay protection devices, automatic devices, and measurement and control devices; 7. Alarm Attribution Relationship: The attribution mapping relationship between secondary functional equipment entities and alarm signals, used to represent the source device, source functional module, and belonging logic unit of various alarm signals;

[0039] The status feature extraction module extracts event time sequence segment features, operational health features, command response features, alarm co-occurrence features, communication link quality features, configuration consistency features, and cross-device linkage features from standardized monitoring time sequence data to form a secondary device status feature vector;

[0040] The model training and management module constructs an adaptive dynamic graph structure model. Based on historical standardized monitoring time-series data, historical secondary equipment state feature vectors, and corresponding historical operating state labels, offline training is performed on the adaptive dynamic graph structure model to generate a trained adaptive dynamic graph structure model. The construction method of the adaptive dynamic graph structure model is as follows: based on graph neural networks, a trainable dynamic adjacency weight mechanism and a cross-time window state propagation mechanism are introduced to optimize the fixed topology and static node representation in the graph neural network. The graph structure is adaptively adjusted by constructing a dynamic adjacency matrix that evolves over time, and the cross-window dependency modeling capability is enhanced by combining time-series embedding recursion and historical state accumulation mechanisms, thus constructing an adaptive dynamic graph structure model with spatiotemporal coupling representation capability. The adaptive dynamic graph structure model includes a dynamic graph neural network for graph structure modeling, a long short-term memory network for time-series modeling, a training inner loop unit during testing, and a variational graph autoencoder network for constructing graph structure constraint loss during offline training.

[0041] The operation status assessment module calls the trained adaptive dynamic graph structure model to perform online reasoning processing on the current secondary equipment status feature vector, and outputs the secondary equipment operation status judgment result. The secondary equipment operation status judgment result includes normal status, abnormal fluctuation status, communication abnormal status, alarm abnormal status, action mismatch status, clock abnormal status, interlocking abnormal status and suspected fault status.

[0042] In this embodiment, a 220kV substation is selected, including 2 220kV main transformer bays, 6 110kV outgoing line bays, and 12 10kV distribution bays; there are a total of 128 secondary devices in the station, including 32 relay protection devices, 28 measurement and control devices, 6 fault recording devices, 4 automatic devices, 2 remote terminal units, 2 communication management units, 18 switches, 16 merging units, 14 smart terminals, and 2 time synchronization devices;

[0043] After processing by fusing dynamic graph neural networks and LSTM, the probability distribution of each device's state is output within this time window. Taking P110-2 as an example: normal state probability: 0.08; abnormal fluctuation state probability: 0.12; communication abnormal state probability: 0.21; alarm abnormal state probability: 0.16; action mismatch state probability: 0.31; clock abnormal state probability: 0.05; interlocking abnormal state probability: 0.04; suspected fault state probability: 0.03. The model threshold is set to 0.25, and the maximum probability is the action mismatch state (0.31). Therefore, P110-2 is determined to be in the "action mismatch state" within this window.

[0044] The anomaly correlation reasoning module combines the secondary equipment object model, correlation graph, and operation status judgment results to perform correlation reasoning on the anomaly propagation relationships across devices, bays, and communication links, generating anomaly correlation reasoning results. Based on the operation status judgment results and anomaly correlation reasoning results, it generates equipment-level, bay-level, and station-level status assessment results, and outputs alarm levels, anomaly types, evidence event sets, suggested verification paths, and handling suggestions to operation and management personnel to support substation operation management and risk control decisions.

[0045] Status assessment results:

[0046] 1. Equipment-level assessment results: Equipment: P110-2 protection device; Operating status: Action mismatch; Anomaly type: Inconsistent protection startup and feedback; Alarm level: Level 2 alarm; Evidence event set: 1) Protection startup signal 3 times; 2) Trip remote signal not corresponding; 3) Network port anomaly 5 times; 4) Message loss rate 2.4%; Suggested verification path: 1) Check the connection from the protection device to the switch port; 2) Check the corresponding port error statistics; 3) Check the status of the optical module; Handling suggestions: 1) Prioritize replacing the abnormal port module; 2) Perform link error test; 3) Strengthen link redundancy switching verification.

[0047] 2. Bay-level assessment results: Bay: 110kV outgoing line section II bay; Overall status results of 4 devices in this bay: 1 device has mismatched operation, 1 device has communication abnormality, and 2 devices have abnormal fluctuations; Assessment result: Bay-level operational risk is moderate; Alarm level: Level II; Abnormal propagation path: Switch port abnormality → Communication packet loss → Protection action mismatch.

[0048] 3. Station-level assessment results: Statistics of 128 devices in the entire station: normal: 113, abnormal fluctuations: 7, communication anomalies: 3, action mismatch: 2, alarm anomalies: 2, other anomalies: 1; station-level health index: 0.91; station-level status assessment: overall operation is stable, but there is a trend of degradation in local communication links; alarm level: Level 3 warning.

[0049] Example 2, based on Example 1, includes the following feature construction content in the state feature extraction module:

[0050] Based on the remote signaling change records, SOE sequence event records and protection action information contained in the standardized monitoring time series data, related events in the same interval are aggregated according to the preset time window to construct an event sequence window; within the event sequence window, the event occurrence order features, adjacent event interval time features, unit time action trigger frequency features, and abnormal event combination pattern features are extracted to generate event time series segment features corresponding to the device.

[0051] Based on telemetry data and device self-test information in standardized monitoring time series data, operational statistical features are constructed within a preset time window. Features of electrical quantity fluctuation amplitude, periodic change, over-limit ratio, self-test anomaly count, and continuous stable running time are extracted to generate operational health features that characterize the stability and health status of equipment operation.

[0052] Based on the remote control issuance records and corresponding status feedback records in the standardized monitoring time series data, a command-response matching relationship is established under a unified time axis. Command execution statistical features are constructed, and command response success rate features, response delay distribution features, non-response count features, and abnormal feedback type features are extracted to generate command response features, which are used to characterize the consistency, timeliness, and abnormal response behavior during the execution of equipment control commands.

[0053] Based on alarm information in standardized monitoring time series data, an alarm statistical model is constructed within a preset time window. Alarm density features, alarm count features per unit time, alarm duration features, and alarm co-occurrence matrix features are extracted to generate alarm co-occurrence features, which are used to characterize the concentration, persistence, and multi-alarm co-occurrence patterns of equipment alarms.

[0054] Based on the communication link status information and network status records in the standardized monitoring time series data, link operation statistical analysis is performed in the time dimension to extract link interruption frequency characteristics, link recovery time characteristics, packet loss rate characteristics, port anomaly count characteristics, and network jitter characteristics, and generate communication link quality characteristics to characterize the stability and reliability level of the device's communication channel.

[0055] Based on the pressure plate status information, setpoint area information and clock synchronization status information in the standardized monitoring time sequence data, and combined with the configuration constraint rules in the equipment object model, configuration status consistency features are constructed, and function activation status identifier features, setpoint area switching record features, pressure plate activation / deactivation consistency features and clock synchronization stability features are extracted to generate configuration consistency features, which are used to characterize whether the current operating configuration of the equipment meets the engineering logic constraints and protection function activation conditions.

[0056] Based on the relationship graph and standardized monitoring time series data, we construct the inter-device collaboration features in the topology and time dimensions, extract the matching features of protection actions and measurement and control feedback, the time difference features of event propagation between devices, the time synchronization dependency consistency features, and the link anomaly linkage features, and generate cross-device linkage features to characterize the collaborative operation behavior and anomaly propagation patterns between different devices.

[0057] Example 3, according to Figure 2This embodiment is based on Embodiment 2. In this embodiment, the process of offline training of the adaptive dynamic graph structure model by the model training and management module includes: constructing a training sample set containing historical standardized monitoring time series data, historical secondary equipment state feature vectors, and historical operating state labels; generating historical spatiotemporal dynamic graph representation sequences corresponding to the training samples based on the training sample set; performing graph augmentation processing on the historical spatiotemporal dynamic graph representation sequences to construct a contrastive loss term, inputting the historical spatiotemporal dynamic graph representation sequences into a variational graph autoencoder network to construct a VGAE loss term, and constructing a classification loss term based on the historical operating state labels; weightedly fusing the contrastive loss term, VGAE loss term, and classification loss term to construct a total loss function; and performing end-to-end backpropagation and joint optimization training based on the total loss function to obtain the trained adaptive dynamic graph structure model.

[0058] In this embodiment, the convergence curves of each sub-loss during the training process of the adaptive dynamic graph structure model are as follows: Figure 2 As shown; Figure 2 In the figure, the horizontal axis represents the "training rounds," indicating the iterative process of the model from round 1 to round 200; the vertical axis represents the "loss value," indicating the magnitude of the corresponding loss function in the current round. The figure shows the changing trends of the contrastive loss, VGAE loss term, and classification loss term during the training process. The contrastive loss term is used to constrain the consistency of the spatiotemporal dynamic graph representation sequence under different augmented views, the VGAE loss term is used to constrain the reconstruction capability of the dynamic graph structure, and the classification loss term is used to constrain the matching relationship between the running state determination result and the true state label.

[0059] Example 4, based on Example 3, describes the process of using a trained adaptive dynamic graph structure model to process the secondary equipment state feature vector and output the secondary equipment operating status determination result. The specific steps include:

[0060] Step S1: Collect, align, and concatenate the current secondary equipment state feature vectors according to a unified time window to obtain the secondary equipment state analysis input tensor. Divide the secondary equipment state analysis input tensor into a time window sequence according to the time dimension. Use the first time window in the time window sequence as the initial graph construction window. Construct an initial adjacency matrix based on the relationship graph and the feature nodes within the first time window. Generate an initial node state representation based on the feature sub-vectors of each feature node within the first time window. The initial adjacency matrix includes object topology adjacency relationships, communication link adjacency relationships, timing dependency adjacency relationships, and configuration constraint adjacency relationships determined by the relationship graph. The initial node state representation is obtained by linearly mapping the feature sub-vectors of each feature node within the first time window.

[0061] Step S2: For the current time window in the time window sequence, if the current time window is the first time window, the initial node state representation is used as the historical state representation of the current time window; if the current time window is a subsequent time window after the first time window, the node state representation output by the previous time window is used as the historical state representation of the current time window; based on the historical state representation of the current time window, a time propagation edge is established between adjacent time windows of the same feature node, and a dynamic adjacency matrix and a current node state representation corresponding to the current time window are constructed; the dynamic adjacency matrix is ​​used to perform message passing and node embedding aggregation operations; the dynamic adjacency matrix corresponding to the current time window is obtained by fusing the initial adjacency matrix, the time adjacency relationship corresponding to the time propagation edge, and the similarity relationship between the current node state representation;

[0062] Step S3: Call the test-time training inner loop unit on the time window sequence to perform a history-dependent dynamic feature hybrid update on the secondary device state analysis input tensor, generating the inner loop updated feature tensor; input the inner loop updated feature tensor into the Long Short-Term Memory (LSTM) network in parallel to generate a temporal embedding matrix, which is used to characterize the state evolution features across time windows; Test-time training inner loop unit: Based on the model having been trained offline, when facing the current real-time monitoring data, temporarily learn the feature relationships within the current window, dynamically correct the input features, and then feed them into the LSTM and dynamic graph neural network for state evaluation;

[0063] Step S4: For the dynamic adjacency matrix corresponding to the current time window, construct an interpretability weight matrix with the same dimension as the dynamic adjacency matrix;

[0064] Step S5: In the dynamic graph neural network, for each time window, the dynamic adjacency matrix, interpretability weight matrix, and temporal embedding matrix are fused to obtain the spatiotemporal graph representation of that time window. After fusing the time windows piece by piece, a spatiotemporal dynamic graph representation sequence corresponding to the secondary device state feature vector is formed. Among them, the interpretability weight matrix and the dynamic adjacency matrix are multiplied element by element to strengthen key nodes and key edges, and the temporal embedding information is injected into the graph representation to introduce single feature time change information.

[0065] Step S6: Input the spatiotemporal dynamic graph representation sequence into the classification output layer of the trained adaptive dynamic graph structure model to obtain the state probability distribution of each preset operating state category, and output the secondary equipment operating state determination result according to the state probability distribution.

[0066] Example 5 differs from Example 4 in that: the inner loop unit trained during testing is invoked on the time window sequence to perform a dynamic feature fusion update based on historical dependencies on the secondary device state analysis input tensor; step S3 differs from Example 3. In this example, step S3 specifically includes the following: a feature update mechanism based on historical state recursion is introduced on the time window sequence to perform a dynamic feature fusion update across time windows on the secondary device state analysis input tensor, generating an updated feature tensor; the updated feature tensor is input into a Long Short-Term Memory (LSTM) network to generate a temporal embedding matrix, which is used to characterize the state evolution features across time windows.

[0067] Example 6 differs from Example 4 in that: the inner loop unit trained during testing is invoked on the time window sequence to perform a dynamic feature fusion update based on historical dependencies on the secondary device state analysis input tensor; step S3 differs from Example 3; step S3 specifically includes the following: a feature update mechanism based on cyclic recursion is introduced on the time window sequence, and the hidden state of the previous time window is gated and fused with the features of the current window to generate an updated feature tensor; the updated feature tensor is input into a Long Short-Term Memory (LSTM) network to generate a temporal embedding matrix, which is used to characterize the state evolution features across time windows.

[0068] Example 7, based on Example 4, specifically includes the following steps in step S3: Training the inner loop unit during testing and generating the updated feature tensor.

[0069] Step S31: For the j-th feature node of the t-th time window in the time window sequence, extract the corresponding feature sub-vector from the secondary device status analysis input tensor, and map it through the query projection matrix, key projection matrix and value projection matrix respectively to obtain the query vector, key vector and value vector that correspond one-to-one with the time window and the feature node.

[0070] Step S32: Construct the TTT inner loop mapping function as the object of fast weight update. The TTT inner loop mapping function consists of a pre-representation function and a linear output weight matrix. Input the key vector into the pre-representation function to obtain the effective key vector, and input the key vector into the TTT inner loop mapping function to obtain the key-induced output.

[0071] The inner loop mapping function of TTT satisfies the following relationship:

[0072] ;

[0073] in, Indicates the time window index. Indicates the feature node index. Indicates the first The first time window The key vector corresponding to each feature node; This indicates the parameters of the prepended function; This indicates that the key vector The effective key vector obtained after inputting the pre-representation function; This represents a linear output weight matrix; This indicates that the preceding representation function is used. and linear output weight matrix The inner loop mapping function of TTT; This indicates that the key vector The key-induced output obtained after inputting the TTT inner loop mapping function is used for self-supervised matching with the value vectors corresponding to the same time window and the same feature node.

[0074] Step S33: Using the value vectors corresponding to the same time window and the same feature node as the self-supervised target value of the key-induced output, a self-supervised inner loop loss function is constructed based on the residual term between the key-induced output and the value vector and the structural consistency constraint term determined by the dynamic adjacency matrix. The self-supervised inner loop loss function is constructed using a linear expandable target function in the form of key-value binding, so that after performing a single-step fast weight update on the TTT inner loop mapping function, its inner loop output can be converted into a linear attention-type incremental expression based on effective key vectors and effective value vectors. The self-supervised inner loop loss function includes a key-value consistency term and a structural consistency constraint term. The key-value consistency term is used to constrain the consistency between the key-induced output and the value vector corresponding to the same time window and the same feature node in the representation space. The structural consistency constraint term is used to maintain the output structural consistency between adjacent feature nodes represented by the dynamic adjacency matrix within the same time window.

[0075] The self-supervised inner loop loss function is defined as:

[0076] ;

[0077] in, Indicates the first The self-supervised inner loop loss function corresponding to each time window; Indicates the feature node index; Indicates and The value vectors that are in the same time window and the same feature node are used as the self-supervised target values ​​of the key-induced output; Represents the weighting coefficients of the structural consistency constraint terms; Indicates the first The dynamic adjacency matrix corresponding to the nth time window is the nth... The feature node and the first Dynamic adjacency weights between feature nodes; the first term is used to constrain the consistency between the key-induced output and the corresponding value vector, and the second term is used to constrain the consistency of the output structure of adjacent feature nodes under dynamic adjacency relationships;

[0078] Step S34: Perform a single-step fast weight update on the TTT inner loop mapping function based on the self-supervised inner loop loss function to obtain the updated inner loop parameter state, and construct an effective value vector based on the residual direction between the value vector and the key-induced output corresponding to the same time window and the same feature node; construct a node increment term based on the effective key vector and the effective value vector, and aggregate the node increment terms corresponding to each feature node in the current time window to obtain the window increment term of the current time window; superimpose the window increment term of the current time window with the preset initial state matrix to obtain the cumulative state matrix corresponding to the current time window; the single-step fast weight update is used to generate the temporary inner loop parameter state corresponding to the current time window, without writing back the global model parameters of the adaptive dynamic graph structure model after training; the node increment term is composed of the outer product of the effective key vector and the effective value vector corresponding to the feature node; the effective key vector and the effective value vector are both derived from the key vector and value vector corresponding to the same time window and the same feature node, and the two are paired one-to-one according to the time window index t and the feature node index j;

[0079] A single-step fast weight update is performed on the TTT inner loop mapping function based on the self-supervised inner loop loss function to obtain the updated inner loop parameter state; the fast parameter update formula is defined as:

[0080] ;

[0081] in, Indicates the first The linear output weight matrix in the TTT inner loop mapping function before the inner loop update is executed in each time window; Indicates the first Each time window is based on a self-supervised inner loop loss function. The linear output weight matrix obtained after performing a fast weight update; Indicates the first Before the inner loop updates are executed within a time window, the parameter set of the pre-representation function is used. Indicates the first The set of parameters for the pre-representation function obtained after performing a fast weight update once within a time window; Indicates the first The learning rate within each time window controls the magnitude of each single-step update; Indicates the joint parameters Find the gradient; Indicates the first The first time window Key vectors corresponding to each feature node The key-induced output obtained after inputting the inner loop mapping function of TTT;

[0082] Step S35: Extract the query vector from the secondary device status analysis input tensor corresponding to the current time window, input the query vector into the pre-representation function to obtain the effective query vector, and obtain the window-level inner loop output vector based on the effective query vector, the updated inner loop parameter status, and the cumulative status matrix corresponding to the current time window;

[0083] Step S36: Perform residual adaptive fusion between the window-level inner loop output vector and the original feature representation corresponding to the current time window to obtain the window-level inner loop updated feature representation, and stack the window-level inner loop updated feature representations in the order of the time windows to form the feature tensor after the inner loop update.

[0084] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An intelligent assessment system for substation operation status based on artificial intelligence, characterized in that, The system includes: The data access module acquires standardized raw monitoring data; The data governance module processes standardized monitoring raw data to generate standardized monitoring time-series data; The equipment modeling module constructs a model of secondary equipment objects and their associated relationship graphs. The status feature extraction module extracts event time sequence segment features, operational health features, command response features, alarm co-occurrence features, communication link quality features, configuration consistency features, and cross-device linkage features from standardized monitoring time sequence data to form a secondary device status feature vector; The model training and management module constructs an adaptive dynamic graph structure model. Based on historical standardized monitoring time-series data, historical secondary equipment status feature vectors, and corresponding historical operating status labels, it performs offline training on the adaptive dynamic graph structure model to generate a trained adaptive dynamic graph structure model. The adaptive dynamic graph structure model includes a dynamic graph neural network, a long short-term memory network, a test-time training inner recurrent unit, and a variational graph autoencoder network. The operational status assessment module calls the trained adaptive dynamic graph structure model to perform online inference processing on the current secondary equipment status feature vector and outputs the operational status judgment result. The anomaly correlation reasoning module combines the secondary equipment object model, correlation graph, and operating status determination results to perform reasoning. The process of calling the trained adaptive dynamic graph structure model to perform online inference processing on the current secondary equipment state feature vector and output the operating status determination result includes the following steps: Step S1: Collect, align and concatenate the current secondary equipment state feature vectors according to a unified time window to obtain the secondary equipment state analysis input tensor, and divide the secondary equipment state analysis input tensor into a time window sequence according to the time dimension; construct a window using the first time window in the time window sequence as the initial graph, and generate the initial node state representation based on the feature sub-vectors of each feature node in the first time window; Step S2: For the current time window in the time window sequence, if the current time window is the first time window, the initial node state representation is used as the historical state representation of the current time window; if the current time window is a subsequent time window after the first time window, the node state representation output by the previous time window is used as the historical state representation of the current time window; based on the historical state representation of the current time window, a time propagation edge is established between adjacent time windows of the same feature node, and a dynamic adjacency matrix corresponding to the current time window is constructed; the dynamic adjacency matrix is ​​used to perform message passing and node embedding aggregation operations. Step S3: Call the inner loop unit trained during testing on the time window sequence to perform dynamic feature fusion update based on historical dependencies on the secondary device status analysis input tensor, generating the feature tensor updated by the inner loop; input the feature tensor updated by the inner loop into the long short-term memory network to generate the temporal embedding matrix; Step S4: For the dynamic adjacency matrix corresponding to the current time window, construct an interpretability weight matrix with the same dimension as the dynamic adjacency matrix; Step S5: In the dynamic graph neural network, for each time window, the dynamic adjacency matrix, interpretability weight matrix and temporal embedding matrix are fused to obtain the spatiotemporal graph representation of that time window. Then, the spatiotemporal graph representations are fused piece by piece according to the time window order to form a sequence of spatiotemporal dynamic graph representations. Step S6: Input the spatiotemporal dynamic graph representation sequence into the classification output layer of the trained adaptive dynamic graph structure model, and output the running state determination result; The model training and management module performs offline training on the adaptive dynamic graph structure model, including: constructing a training sample set containing historical standardized monitoring time-series data, historical secondary equipment state feature vectors, and historical operating state labels; generating historical spatiotemporal dynamic graph representation sequences corresponding to the training samples based on the training sample set; performing graph augmentation processing on the historical spatiotemporal dynamic graph representation sequences to construct a contrastive loss term, inputting the historical spatiotemporal dynamic graph representation sequences into a variational graph autoencoder network to construct a VGAE loss term, and constructing a classification loss term based on the historical operating state labels; weightedly fusing the contrastive loss term, VGAE loss term, and classification loss term to construct a total loss function; and performing end-to-end backpropagation and joint optimization training based on the total loss function to obtain the trained adaptive dynamic graph structure model.

2. The intelligent substation operation status assessment system based on artificial intelligence according to claim 1, characterized in that: The state feature extraction module is specifically used for: Based on the remote signaling change records, SOE sequence event records, and protection action information contained in the standardized monitoring time series data, an event sequence window is constructed. Within the event sequence window, the event occurrence sequence features, adjacent event interval time features, unit time action trigger frequency features, and abnormal event combination pattern features are extracted to generate event time series segment features. Based on telemetry data and device self-test information in standardized monitoring time series data, characteristics of electrical quantity fluctuation amplitude, periodic change, over-limit ratio, self-test anomaly count, and continuous stable running time are extracted to generate operational health characteristics. Based on the remote control issuance records and corresponding status feedback records in the standardized monitoring time series data, the command response success rate characteristics, response delay distribution characteristics, non-response count characteristics, and abnormal feedback type characteristics are extracted to generate command response characteristics; Based on alarm information in standardized monitoring time-series data, alarm density features, alarm count features per unit time, alarm duration features, and alarm co-occurrence matrix features are extracted to generate alarm co-occurrence features. Based on the communication link status information and network status records in the standardized monitoring time series data, perform link operation statistical analysis, extract link interruption frequency characteristics, link recovery time characteristics, packet loss rate characteristics, port anomaly count characteristics, and network jitter characteristics, and generate communication link quality characteristics. Based on the pressure plate status information, setpoint area information and clock synchronization status information in the standardized monitoring time sequence data, extract the function activation status identifier features, setpoint area switching record features, pressure plate activation / deactivation consistency features and clock synchronization stability features to generate configuration consistency features. Based on the correlation graph and standardized monitoring time sequence data, the matching features of protection actions and measurement and control feedback, the time difference features of event propagation between devices, the consistency features of time synchronization dependency, and the linkage features of link anomalies are extracted to generate cross-device linkage features.

3. The intelligent substation operation status assessment system based on artificial intelligence according to claim 1, characterized in that: The process of training the inner loop unit during testing and generating the updated feature tensor includes the following steps: Step S31: For the j-th feature node of the t-th time window in the time window sequence, extract the corresponding feature sub-vector from the secondary device status analysis input tensor, and map it through the query projection matrix, key projection matrix and value projection matrix respectively to obtain the query vector, key vector and value vector that correspond one-to-one with the time window and the feature node. Step S32: Construct the TTT inner loop mapping function as the object of fast weight update. The TTT inner loop mapping function consists of a pre-representation function and a linear output weight matrix. Input the key vector into the pre-representation function to obtain the effective key vector, and input the key vector into the TTT inner loop mapping function to obtain the key-induced output. Step S33: Using the value vectors corresponding to the same time window and the same feature node as the self-supervised target values ​​of the key-induced output, construct a self-supervised inner loop loss function based on the residual term between the key-induced output and the value vector and the structural consistency constraint term determined by the dynamic adjacency matrix; Step S34: Perform a single-step fast weight update on the TTT inner loop mapping function based on the self-supervised inner loop loss function to obtain the updated inner loop parameter state, and construct an effective value vector based on the residual direction between the value vector and the key-induced output corresponding to the same time window and the same feature node; construct a node increment term based on the effective key vector and the effective value vector, and aggregate the node increment terms corresponding to each feature node in the current time window to obtain the window increment term of the current time window; superimpose the window increment term of the current time window with the preset initial state matrix to obtain the cumulative state matrix corresponding to the current time window; Step S35: Input the query vector into the pre-representation function to obtain the effective query vector, and obtain the window-level inner loop output vector based on the effective query vector, the updated inner loop parameter state, and the cumulative state matrix corresponding to the current time window; Step S36: Perform residual adaptive fusion between the window-level inner loop output vector and the original feature representation corresponding to the current time window to obtain the window-level inner loop updated feature representation, and stack the window-level inner loop updated feature representations in the order of the time windows to form the feature tensor after the inner loop update.

4. The intelligent substation operation status assessment system based on artificial intelligence according to claim 3, characterized in that: The node increment term is composed of the outer product of the effective key vector and the effective value vector corresponding to the feature node; the effective key vector and the effective value vector are both derived from the key vector and value vector corresponding to the same feature node in the same time window, and the two are paired one by one according to the time window index t and the feature node index j.

Images