Power grid operation risk intelligent identification method and system

Abstract

The application provides a power grid operation risk intelligent identification method and system, relates to the power grid technical field, comprises the following steps: constructing a knowledge graph based on a predefined ontology model; performing multi-hop reasoning based on a time sequence causal reasoning rule, identifying a risk correlation chain, and generating a risk reasoning subgraph; evaluating a risk level; and generating a device regulation instruction and a topology reconstruction scheme through reinforcement learning. The application can realize automatic identification, evaluation and intelligent decision of power grid risks, improve power grid operation safety and reliability, reduce manual intervention, and enhance the ability of the system to resist complex faults.

Description

Technical Field

[0001] This invention relates to power grid technology, and more particularly to a method and system for intelligent identification of power grid operation risks. Background Technology

[0002] As a critical national infrastructure, the safe and stable operation of the power system is crucial for ensuring economic and social development. With the continuous expansion and increasing complexity of the power grid, the operational risks it faces are becoming increasingly complex and volatile. Traditional methods for identifying and managing power grid risks mainly rely on expert experience and static rules, which are insufficient to address the dynamically changing risk scenarios in modern large-scale and complex power grid systems. In recent years, with the rapid development of artificial intelligence technology, risk identification methods based on knowledge graphs and intelligent reasoning have been gradually applied to the field of power grid security, providing a new technological approach for power grid risk management.

[0003] Traditional risk identification methods lack a deep understanding of the complex relationship between power grid topology and equipment status, making it difficult to effectively identify complex fault propagation paths caused by multi-device interactions. In particular, they struggle to accurately capture the dynamic risk evolution process under system disturbances. Most existing methods employ static threshold judgment mechanisms for risk assessment, failing to fully consider the temporal characteristics and causal relationships of power grid operation. This results in delayed risk warnings and difficulty in identifying potential risks in advance. Existing risk management systems generally lack adaptive learning capabilities, making it impossible to optimize and adjust risk identification models based on historical decision-making results. The system's decision support often relies on preset rules, making it difficult to cope with complex and ever-changing abnormal power grid conditions. Summary of the Invention

[0004] The present invention provides a method and system for intelligent identification of power grid operation risks, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for intelligent identification of power grid operation risks, comprising:

[0006] Acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph to identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges;

[0007] Based on the risk source nodes and impact range information in the risk reasoning subgraph, combined with the power grid operation constraints and safety margin requirements, the risk level is assessed to obtain the risk level assessment result.

[0008] Based on the risk level assessment results and the current operating status of the power grid, the decision strategy generation module based on reinforcement learning outputs a set of autonomous decision-making instructions for equipment control and topology reconfiguration schemes. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the set of autonomous decision-making instructions, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

[0009] Multi-hop reasoning based on temporal causal reasoning rules is performed on the entity nodes and relational edges in the knowledge graph to identify fault propagation paths and risk association chains, generating a risk reasoning subgraph containing risk source nodes, propagation paths, and impact scope, including:

[0010] In the knowledge graph, device entity nodes whose state attributes have changed abnormally are detected, and the detected device entity nodes, their timestamps, and state change features are encapsulated to form initial inference anchors.

[0011] Starting from the initial inference anchor point, the system traverses outward along the association edges in the knowledge graph. It then verifies the causal transitivity and temporal sequence of the subsequent device entity nodes connected by the association edges using temporal causal inference rules. Finally, it combines the verified subsequent device entity nodes and their association edges with the preceding nodes to form a propagation path segment.

[0012] The subsequent device entity nodes in the propagation path segment are updated to new inference anchor points. The aforementioned traversal and verification process is repeated. Propagation path segments are continuously superimposed until the subsequent device entity nodes connected by the association edge cannot pass the verification of the temporal causal inference rule. The superimposed propagation path segments are connected to form multiple complete fault propagation paths starting from the initial inference anchor point.

[0013] Based on the device entity nodes and associated edges contained in the multiple complete fault propagation paths, a risk reasoning subgraph is constructed.

[0014] The causal transitivity and temporal sequence of the subsequent device entity nodes connected by the associated edges are verified using temporal causal reasoning rules. The verified subsequent device entity nodes and their associated edges with the preceding nodes are combined to form a propagation path segment, including:

[0015] The state change timestamps and abnormal state features of the preceding nodes are obtained from the inference anchor point, and the state change timestamps and abnormal state features of the connected subsequent device entity nodes are obtained through the association edges based on the preceding nodes.

[0016] The time interval is calculated based on the state change timestamps of the preceding node and the following device entity node. The time interval is compared with the physical propagation delay of the associated edge to complete the temporal sequence verification. At the same time, the abnormal state characteristics of the preceding node and the following device entity node are input into a preset causal transit rule base for matching to complete the causal transitivity verification.

[0017] For node pairs that have completed the sequentiality verification and causal transitivity verification, the subsequent device entity node is bound to its corresponding association edge. During the binding process, the time interval value and the causal rule matching result are written into the attribute field of the association edge, ultimately forming a propagation path segment containing complete verification information.

[0018] Based on the risk source nodes and impact range information in the risk reasoning subgraph, and combined with the power grid operation constraints and safety margin requirements, the risk level is assessed, and the risk level assessment results include:

[0019] The device importance of the risk source node and the topology depth of the fault propagation path are extracted from the risk reasoning subgraph. The device importance and the topology depth are adaptively coupled and calculated, and an electrical distance attenuation factor is introduced to form a risk propagation intensity feature that includes spatial attenuation characteristics.

[0020] The affected equipment entity nodes in the scope of impact information and the operating nodes in the power grid operation constraints are identified through multi-dimensional intersection. For the node group in the intersection, the dynamic deviation between its current operating status and the hierarchical safety margin requirements is calculated. At the same time, the coupling correlation strength between nodes is considered to construct a safety sensitivity feature with group characteristics.

[0021] The risk propagation intensity feature and the security sensitivity feature are input into the nonlinear fusion module. During the fusion process, the fusion weights are dynamically adjusted according to the hop depth of the fault propagation path and the node group features in the risk reasoning subgraph to generate a risk level assessment result that reflects the overall risk situation of the system.

[0022] The risk propagation intensity feature and the security sensitivity feature are input into the nonlinear fusion module. During the fusion process, the fusion weights are dynamically adjusted based on the hop depth of the fault propagation path and the node group features in the risk reasoning subgraph, including:

[0023] The risk reasoning subgraph traverses all fault propagation paths and obtains the hop depth of each path. Based on the hop depth, the propagation extension index is calculated. At the same time, the voltage level distribution and topology connection density of the affected equipment entity nodes are read from the risk reasoning subgraph. Based on the voltage level distribution and topology connection density, node group features are constructed. The propagation extension index and the node group features are fed into the weight generation function and the propagation weight factor and sensitivity weight factor are calculated.

[0024] The risk propagation intensity feature is weighted by the propagation weight factor to obtain a weighted risk propagation component, and the security sensitivity feature is weighted by the sensitivity weight factor to obtain a weighted security sensitivity component. In the nonlinear fusion module, a nonlinear transformation is applied to the weighted risk propagation component and the weighted security sensitivity component, and they are accumulated to generate a comprehensive risk feature.

[0025] Based on the risk level assessment results and the current operating status of the power grid, the decision-making instruction set for equipment control commands and topology reconfiguration schemes, generated by the reinforcement learning-based decision-making strategy generation module, includes:

[0026] The risk level identifier and the set of affected equipment nodes are read from the risk level assessment results. At the same time, the power flow distribution data and equipment operating parameters are read from the current power grid operating status. Feature encoding is performed on the risk level identifier and the set of affected equipment nodes to obtain a risk state representation. Feature encoding is performed on the power flow distribution data and the equipment operating parameters to obtain an operating state representation. The risk state representation and the operating state representation are concatenated to form a decision state vector. The decision state vector is then sent to the reinforcement learning-based decision strategy generation module.

[0027] The decision state vector is processed by feature extraction and strategy mapping according to the decision strategy generation module, and outputs a device control action sequence and a topology reconstruction action sequence. Based on the device control action sequence, control instructions for the control target device are generated, and based on the topology reconstruction action sequence, a reconstruction scheme for the network topology connection is generated. The control instructions and the reconstruction scheme are combined to form a complete set of autonomous decision instructions.

[0028] The decision strategy generation module maps historical decision execution results to the weight parameter space of the time-series causal reasoning rule for adjustment, executes the autonomous decision instruction set, and updates the power grid operation status data after decision execution to the corresponding node attributes of the knowledge graph, including:

[0029] The feature quantities of historical decision execution results are mapped to the weight parameter space of the time-series causal reasoning rule, and the weight parameters are adjusted accordingly based on the mapping results; based on the adjusted weight parameters, an autonomous decision instruction set including equipment control instructions and topology reconfiguration schemes is executed; the power grid operation status data after the decision instructions are executed is obtained, and the power grid operation status data is updated and written into the attribute fields of the corresponding nodes in the knowledge graph.

[0030] A second aspect of the present invention provides a power grid operation risk intelligent identification system, comprising:

[0031] The first module is used to acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph, identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges.

[0032] The second module is used to assess the risk level based on the risk source nodes and impact range information in the risk reasoning subgraph, combined with the power grid operation constraints and safety margin requirements, and obtain the risk level assessment result.

[0033] The third module is used to output an autonomous decision-making instruction set of equipment control instructions and topology reconfiguration schemes based on the risk level assessment results and the current operating status of the power grid through a reinforcement learning-based decision strategy generation module. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the autonomous decision-making instruction set, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

[0034] A third aspect of the present invention provides an electronic device, comprising:

[0035] processor;

[0036] Memory used to store processor-executable instructions;

[0037] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0038] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0039] The beneficial effects of this application are as follows:

[0040] By constructing a knowledge graph based on a predefined ontology model and combining it with temporal causal reasoning rules for multi-hop reasoning, it is possible to effectively identify fault propagation paths and risk association chains. Compared with traditional methods, it can more comprehensively grasp the dynamic propagation characteristics of power grid risks and improve the accuracy and systematicness of risk identification.

[0041] This invention generates a risk reasoning subgraph that includes risk source nodes, propagation paths, and impact ranges, making risk assessment more intuitive and accurate. This allows maintenance personnel to quickly grasp the risk situation and take targeted measures. By combining power grid operation constraints and safety margin requirements to assess risk levels, the risk assessment better aligns with actual power grid operation needs, improving the practicality and reliability of the assessment results.

[0042] By using a reinforcement learning-based decision-making strategy generation module, a closed-loop management system from risk assessment to autonomous decision-making is achieved. Furthermore, the system can feed back historical decision execution results to the weight parameter space of the temporal causal reasoning rules for adaptive adjustment, thereby continuously optimizing and improving risk identification and decision-making capabilities. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the intelligent identification method for power grid operation risks according to an embodiment of the present invention;

[0044] Figure 2 This is a dynamic flowchart of power grid system risk level assessment and weight optimization according to an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0047] Figure 1 This is a flowchart illustrating the intelligent identification method for power grid operation risks according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0048] Acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph to identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges;

[0049] Based on the risk source nodes and impact range information in the risk reasoning subgraph, combined with the power grid operation constraints and safety margin requirements, the risk level is assessed to obtain the risk level assessment result.

[0050] Based on the risk level assessment results and the current operating status of the power grid, the decision strategy generation module based on reinforcement learning outputs a set of autonomous decision-making instructions for equipment control and topology reconfiguration schemes. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the set of autonomous decision-making instructions, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

[0051] In one optional implementation, multi-hop reasoning based on temporal causal reasoning rules is performed on the entity nodes and association edges in the knowledge graph to identify fault propagation paths and risk association chains, generating a risk reasoning subgraph containing risk source nodes, propagation paths, and impact ranges, including:

[0052] In the knowledge graph, device entity nodes whose state attributes have changed abnormally are detected, and the detected device entity nodes, their timestamps, and state change features are encapsulated to form initial inference anchors.

[0053] Starting from the initial inference anchor point, the system traverses outward along the association edges in the knowledge graph. It then verifies the causal transitivity and temporal sequence of the subsequent device entity nodes connected by the association edges using temporal causal inference rules. Finally, it combines the verified subsequent device entity nodes and their association edges with the preceding nodes to form a propagation path segment.

[0054] The subsequent device entity nodes in the propagation path segment are updated to new inference anchor points. The aforementioned traversal and verification process is repeated. Propagation path segments are continuously superimposed until the subsequent device entity nodes connected by the association edge cannot pass the verification of the temporal causal inference rule. The superimposed propagation path segments are connected to form multiple complete fault propagation paths starting from the initial inference anchor point.

[0055] Based on the device entity nodes and associated edges contained in the multiple complete fault propagation paths, a risk reasoning subgraph is constructed.

[0056] Anomaly detection of entity node state attributes in the knowledge graph is achieved through a multi-layered threshold system. The state attributes of device entity nodes include physical parameters such as temperature, pressure, vibration, and current, as well as enumerated values ​​of operating states. The anomaly detection module employs a sliding window mechanism with a window length of 300 seconds, performing a detection calculation every 30 seconds. For numerical state attributes, anomaly determination is based on statistical deviation analysis; an anomaly is triggered when the deviation between the current value and the historical mean exceeds 2.5 times the standard deviation. For enumerated state attributes, anomaly determination is based on the validity verification of state transitions; anomaly is triggered when there is an unexpected state transition or the duration of the state exceeds the normal range. The detected abnormal device entity node, along with its precise timestamp, values ​​before and after the state change, the magnitude of the change, and the rate of change, are encapsulated into an initial inference anchor data structure. The initial inference anchor data structure includes a device identifier, an anomaly trigger timestamp accurate to milliseconds, a state attribute name, a baseline value before the anomaly, a current value after the anomaly, a quantified value of the deviation degree, and an anomaly type flag.

[0057] The traversal process, starting from the initial inference anchor point and expanding from there, employs a breadth-first search algorithm with a depth limit of 8 hops to prevent path explosion. The types of relational edges include physical connections, signal transmission, control dependencies, power supply relationships, and data flow, each configured with different propagation delay parameters and attenuation coefficients. The temporal causal inference rule verification consists of two stages: causal transitivity verification and temporal sequence verification. Causal transitivity verification checks the directionality and propagation mechanism of relational edges, verifying whether the abnormal state of a preceding node has a physical or logical basis for propagation to subsequent nodes. Temporal sequence verification checks the order of timestamps, requiring that the abnormal time of a preceding node is earlier than that of a subsequent node, with the time difference falling within the preset propagation delay range of the relational edge. The delay range is dynamically determined based on the relation type: 0.1 to 10 seconds for physical connections, 0.01 to 1 second for signal transmission relationships, and 1 to 60 seconds for control dependencies.

[0058] The verification process for successor device entity nodes requires simultaneously satisfying both causal transitivity and temporal sequence to pass verification. The propagation path segment is formed by combining the association edges between a verified successor device entity node and its predecessor node. The path segment data structure records the starting node identifier, ending node identifier, association edge type, measured propagation delay, confidence score, and propagation attenuation coefficient. The confidence score is calculated based on temporal matching accuracy and causal relationship strength, ranging from 0 to 1. Temporal matching accuracy is quantified by the deviation between the actual and expected delays, while causal relationship strength is quantified by the weights of the association edges and the historical propagation success rate.

[0059] When a successor device entity node in a propagation path segment is updated to a new inference anchor point, it needs to inherit the accumulated information of the preceding propagation path, including accumulated propagation delay, accumulated attenuation coefficient, and total path confidence. The repeated traversal and verification process adopts a recursive depth-first strategy, prioritizing the successor node with the highest confidence score for inference continuation during each expansion. When a successor device entity node connected by an association edge cannot be verified by the temporal causal inference rules, the current propagation path terminates. Termination conditions include temporal mismatch, missing causal relationship, accumulated attenuation coefficient below the 0.3 threshold, propagation path length exceeding the maximum hop limit, and the successor node not having an abnormal state.

[0060] In the process of superimposing and connecting propagation path segments to form a complete fault propagation path, it is necessary to handle path branching and merging. Path branching occurs when a single device node simultaneously propagates an anomaly to multiple successor nodes, and each branch independently calculates its propagation confidence and attenuation coefficient. Path merging occurs when anomalies from multiple preceding nodes simultaneously affect the same successor node. A weighted fusion strategy is used to calculate the overall impact intensity, with weight allocation based on the confidence score and arrival time order of each preceding path. The data structure of the complete fault propagation path includes a unique path identifier, a sequence of path nodes, a sequence of path relationship edges, total propagation delay, average confidence, number of affected devices, and path classification label.

[0061] The risk reasoning subgraph is constructed by reorganizing the graph structure based on the device entity nodes and related edges contained in multiple complete fault propagation paths. The subgraph construction process first extracts the set of nodes and edges involved in all propagation paths, and then filters them based on node importance and edge weight. Node importance assessment considers the node's centrality in the propagation network, historical fault frequency, and business impact level; nodes with an importance score below 0.4 are filtered out. Edge weight assessment considers the physical strength of the relationship, propagation success rate, and propagation delay stability; edges with a weight score below 0.3 are filtered out.

[0062] The risk reasoning subgraph is represented using a directed weighted graph data structure. Node attributes include basic device information, current status, anomaly characteristics, importance score, and business tag. Edge attributes include relationship type, propagation parameters, weight score, and historical statistics. The subgraph also includes global attributes, recording the subgraph creation time, number of involved paths, risk level assessment, expected impact range, and recommended handling strategy. The risk level assessment is based on a comprehensive calculation of the number of affected devices, the degree of business criticality, and the propagation speed, and is divided into three levels: low risk, medium risk, and high risk. The corresponding quantitative thresholds are: fewer than 5 affected devices and no critical business, 5 to 20 affected devices or involving some critical business, and more than 20 affected devices or involving core critical business.

[0063] In one optional implementation, the subsequent device entity nodes connected by the association edges are subjected to causal transitivity and temporal sequence verification using temporal causal reasoning rules. The verified subsequent device entity nodes and their association edges with the preceding nodes are combined to form a propagation path segment, including:

[0064] The state change timestamps and abnormal state features of the preceding nodes are obtained from the inference anchor point, and the state change timestamps and abnormal state features of the connected subsequent device entity nodes are obtained through the association edges based on the preceding nodes.

[0065] The time interval is calculated based on the state change timestamps of the preceding node and the following device entity node. The time interval is compared with the physical propagation delay of the associated edge to complete the temporal sequence verification. At the same time, the abnormal state characteristics of the preceding node and the following device entity node are input into a preset causal transit rule base for matching to complete the causal transitivity verification.

[0066] For node pairs that have completed the sequentiality verification and causal transitivity verification, the subsequent device entity node is bound to its corresponding association edge. During the binding process, the time interval value and the causal rule matching result are written into the attribute field of the association edge, ultimately forming a propagation path segment containing complete verification information.

[0067] The inference anchor obtains the state change timestamps and abnormal state features of preceding nodes through the node attribute access interface. The state change timestamps are stored using 64-bit integers with microsecond-level precision. The abnormal state feature data structure includes fields such as anomaly type identifier, state change value, change direction enumeration, and anomaly severity. The anomaly type identifier uses string encoding to distinguish different types such as temperature anomaly, pressure anomaly, and current anomaly. The state change value records the specific deviation magnitude. The change direction enumeration includes three values: rising, falling, and fluctuating. The anomaly severity uses integer values ​​from 1 to 5 to represent increasing severity.

[0068] The implementation of obtaining successor device entity nodes based on the associated edges of the preceding node adopts an adjacency node query algorithm. It obtains a list of all directly connected successor device entity nodes by traversing the outgoing edge set of the preceding node. The associated edge data structure records the edge type, physical propagation delay parameter, and propagation direction information. The edge type includes enumerated values ​​such as electrical connection, mechanical coupling, signal transmission, and control association. The physical propagation delay parameter is defined as the time interval composed of the minimum and maximum propagation delays. Obtaining the state change timestamps and abnormal state characteristics of the successor device entity nodes requires querying the state history within a time window. The starting point of the time window is set to the moment the preceding node's abnormality occurs, and the ending point is set to the starting point plus twice the maximum propagation delay of the associated edges.

[0069] The time interval calculation process subtracts the state change timestamp of the preceding node from the state change timestamp of the subsequent device entity node to obtain the time difference value. The calculation result maintains microsecond-level accuracy and handles negative values. The comparison and verification of the time interval and the physical propagation delay of the associated edge uses interval inclusion judgment. When the calculated time interval value falls within the range of the minimum and maximum values ​​of the physical propagation delay, the timing sequence verification is marked as passed. The physical propagation delay interval is preset according to the associated edge type: the delay interval for electrical connection type is 0.1 ms to 50 ms, the delay interval for mechanical coupling type is 1 ms to 200 ms, the delay interval for signal transmission type is 0.01 ms to 10 ms, and the delay interval for control association type is 10 ms to 2000 ms.

[0070] The abnormal state features of the preceding node and the subsequent device entity node are input into a pre-defined causal propagation rule base for matching and verification. This is achieved through a rule matching engine. The causal propagation rule base is stored in a rule table structure, with each rule defining the preceding abnormal condition, the subsequent abnormal condition, and the matching confidence weight. The rule matching process combines the abnormal state features of the preceding node and the abnormal state features of the subsequent device entity node to form a matching pattern, and then searches the rule base for causal propagation rules that match this pattern. The matching algorithm supports both exact matching and fuzzy matching modes. Exact matching requires that the abnormal state features completely correspond to the rule conditions, while fuzzy matching allows numerical features to meet the matching conditions within a set tolerance range.

[0071] The causal propagation rule base contains various typical equipment fault propagation modes. Temperature conduction rules define the heat transfer mode from high-temperature equipment to adjacent equipment; pressure transmission rules define the propagation mode of pressure anomalies along the pipeline system; current anomaly rules define the cascading propagation mode of electrical faults in the power supply network; and control failure rules define the equipment response failure mode caused by abnormal control commands. Each rule includes elements such as a precondition matching template, a subsequent condition matching template, a time delay constraint, and a matching confidence calculation method. The matching confidence is calculated based on the similarity of anomaly features, the accuracy of time-series matching, and the historical accuracy of the rule, and its value is a floating-point number ranging from 0 to 1.

[0072] Node pairs that have completed both temporal sequence verification and causal transitivity verification are marked as verified. The verification result data structure records information such as verification status, time interval value, matching rule identifier, and overall confidence level. The binding operation between subsequent device entity nodes and associated edges is achieved by updating the extended attribute fields of the associated edges, writing the time interval value to the measured delay field, and writing the causal rule matching result to the matching rule field. The time interval value is stored as a long integer in microseconds, and the causal rule matching result is stored as a combination of a rule identifier string and a floating-point number representing the matching confidence level.

[0073] The attribute field updates during the binding process employ atomic operations to ensure data consistency. The extended attributes of the associated edges include fields such as measured propagation delay, matching rule identifier, matching confidence, verification timestamp, and verification status flag. The measured propagation delay records the actual time interval calculated for this verification; the matching rule identifier records the causal propagation rule number that passed verification; the matching confidence records the degree of confidence in the rule matching; the verification timestamp records the execution time of the verification operation; and the verification status flag indicates whether the edge has passed the complete verification process.

[0074] The final formation of the propagation path segment integrates the verified successor device entity nodes, corresponding associated edges, and complete verification information into a path segment data object. The path segment data structure includes core fields such as start node reference, end node reference, connecting edge reference, measured propagation delay, causal matching result, and path segment confidence. The path segment confidence is calculated as a weighted average of the temporal verification confidence and the causal verification confidence, with a weight allocation of 0.4 for temporal verification and 0.6 for causal verification, reflecting the importance of causal relationships in fault propagation analysis.

[0075] In one optional implementation, based on the risk source nodes and impact range information in the risk reasoning subgraph, and combined with power grid operation constraints and safety margin requirements, the risk level is assessed, and the risk level assessment result includes:

[0076] The device importance of the risk source node and the topology depth of the fault propagation path are extracted from the risk reasoning subgraph. The device importance and the topology depth are adaptively coupled and calculated, and an electrical distance attenuation factor is introduced to form a risk propagation intensity feature that includes spatial attenuation characteristics.

[0077] The affected equipment entity nodes in the scope of impact information and the operating nodes in the power grid operation constraints are identified through multi-dimensional intersection. For the node group in the intersection, the dynamic deviation between its current operating status and the hierarchical safety margin requirements is calculated. At the same time, the coupling correlation strength between nodes is considered to construct a safety sensitivity feature with group characteristics.

[0078] The risk propagation intensity feature and the security sensitivity feature are input into the nonlinear fusion module. During the fusion process, the fusion weights are dynamically adjusted according to the hop depth of the fault propagation path and the node group features in the risk reasoning subgraph to generate a risk level assessment result that reflects the overall risk situation of the system.

[0079] like Figure 2 As shown, the method includes:

[0080] The risk reasoning subgraph extracts the equipment importance of risk source nodes and the topological depth of fault propagation paths. Equipment importance is quantified by a weighted combination of the node's betweenness centrality and degree centrality in the network, specifically representing the node's criticality in the entire power network. For example, the importance value for equipment in the main substation is as high as 0.85-0.95, while for equipment in the terminal distribution lines, the importance value is only 0.2-0.3. Meanwhile, topological depth is represented by the number of hops of the shortest path from the risk source node to the target node; with each increase in topological distance, the impact decreases accordingly.

[0081] An adaptive coupling calculation is performed between device importance and topology depth, incorporating an electrical distance attenuation factor to form a risk propagation strength characteristic that includes spatial attenuation. The adaptive coupling calculation employs a nonlinear function mapping, ensuring that high-importance devices maintain high risk propagation strength at shallower topology depths, while the risk propagation strength decreases exponentially with increasing topology depth. The electrical distance attenuation factor considers the impedance characteristics of the electrical network and is typically expressed in exponential form e. (-λd) The expression represents the distance between two substations, where d is the electrical distance and λ is the attenuation coefficient, which can range from 0.1 to 0.5 depending on the voltage level. In practical applications, when considering the impact of a fault in a substation in a 500kV backbone network on adjacent substations, a smaller attenuation coefficient (such as 0.1) is used to reflect the wide range of impact in high-voltage networks.

[0082] The process involves multi-dimensional intersection identification of affected equipment entity nodes in the impact scope information with operating nodes in the power grid operation constraints. This multi-dimensional intersection identification process first constructs a set of equipment entity nodes and a set of operating nodes. Then, an attribute matching algorithm identifies nodes that coexist in both sets, forming an intersection node group. For example, when the risk source is a fault in a 500kV substation, the affected equipment includes multiple transmission lines connecting the substation and a downstream 220kV substation. The operation constraints involve power flow limits and voltage stability requirements for these lines. Intersection identification can determine the set of equipment nodes requiring key attention.

[0083] For the node group in the intersection set, the dynamic deviation between its current operating state and the hierarchical safety margin requirements is calculated. The hierarchical safety margin requirements are set with different margin thresholds according to different power grid operating levels (such as normal, early warning, alarm, and emergency). For example, for power flow, the margin is required to be no less than 30% in the normal state, while it can be reduced to 10% in the emergency state. The dynamic deviation is calculated using a normalized distance metric, which normalizes the difference between the current operating parameters and the corresponding safety margin threshold to obtain a deviation value between 0 and 1.

[0084] Simultaneously, a security sensitivity feature with group characteristics is constructed considering the coupling strength between nodes. The coupling strength is represented by a correlation matrix, reflecting the degree of interdependence of power flows between nodes; a higher correlation value indicates stronger coupling. The group characteristic employs a community detection algorithm to identify closely related subgroups from the intersection nodes, and assigns different weights to each subgroup to represent the degree of impact of different node clusters on system security.

[0085] Risk propagation intensity characteristics and safety sensitivity characteristics are input into a nonlinear fusion module. The fusion process employs the Choquet fuzzy integral method, which can handle the interactions between features. The fusion weights are dynamically adjusted based on the hop depth of the fault propagation path and the node group characteristics in the risk reasoning subgraph, assigning higher weights to smaller hop counts and larger group sizes. For example, a directly connected one-hop node has a fusion weight of 0.7-0.9, while nodes with three or more hops have a weight reduced to 0.1-0.3. Finally, the fusion result is converted into a risk level assessment score through a nonlinear mapping function, generally divided into four levels: low risk (0-0.3), medium risk (0.3-0.6), high risk (0.6-0.8), and severe risk (0.8-1.0).

[0086] In a practical application scenario, when a circuit breaker fault is detected at a 220kV substation A, the circuit breaker is first identified as the risk source node, with an equipment importance of 0.75. Through risk reasoning subgraph analysis, it is found that the fault affects three connected 110kV lines and two 110kV substations, constituting the impact area. Intersection analysis of these devices with current operating constraints identifies one 110kV line operating under high load, with a current load rate of 85% and a safety margin requirement of 70%, indicating a high degree of dynamic deviation. Considering the topology depth of the fault propagation path is 2 (requiring two hops from the circuit breaker to the 110kV line) and introducing an electrical distance attenuation factor of 0.2, the calculated risk propagation intensity is 0.65. Finally, through nonlinear fusion calculation, the risk level of this risk event is determined to be 0.72, belonging to the high-risk level. Maintenance personnel need to prioritize handling the circuit breaker fault and closely monitor the operating status of the relevant 110kV lines.

[0087] The risk assessment results not only reflect the direct impact of equipment failures, but also consider the dynamic relationship between system operating status and safety margin, providing power grid operators with a comprehensive risk situation assessment, which helps to optimize operation and maintenance decisions and resource allocation.

[0088] In one optional implementation, the risk propagation intensity feature and the security sensitivity feature are input into a nonlinear fusion module. During the fusion process, the fusion weights are dynamically adjusted based on the hop depth of the fault propagation path and the node group characteristics in the risk reasoning subgraph, including:

[0089] The risk reasoning subgraph traverses all fault propagation paths and obtains the hop depth of each path. Based on the hop depth, the propagation extension index is calculated. At the same time, the voltage level distribution and topology connection density of the affected equipment entity nodes are read from the risk reasoning subgraph. Based on the voltage level distribution and topology connection density, node group features are constructed. The propagation extension index and the node group features are fed into the weight generation function and the propagation weight factor and sensitivity weight factor are calculated.

[0090] The risk propagation intensity feature is weighted by the propagation weight factor to obtain a weighted risk propagation component, and the security sensitivity feature is weighted by the sensitivity weight factor to obtain a weighted security sensitivity component. In the nonlinear fusion module, a nonlinear transformation is applied to the weighted risk propagation component and the weighted security sensitivity component, and they are accumulated to generate a comprehensive risk feature.

[0091] The risk reasoning subgraph is a directed graph structure that represents how a fault propagates from one device to another in a power system. Each path represents a fault propagation chain. The hop depth refers to the number of intermediate nodes a fault passes through to propagate from the source node to the target node. For example, if a fault propagates from transformer A to transformer B, then to line C, and finally to bus D, the hop depth of this path is 3.

[0092] The propagation spread index is calculated based on the acquired hop count depth. This index reflects the breadth and depth of fault propagation in the network and can be quantified using the statistical distribution characteristics of the hop count depth. Specifically, the implementation involves first calculating the average hop count (avg) across all paths. hop and maximum value max hop Then, the propagation extension index ED=(avg) is calculated using the normalization formula. hop / max hop )×(1-e (-max_hop / 5) ), where e is the base of the natural logarithm. The value of this index ranges from 0 to 1, and the wider the fault propagation range and the longer the path, the closer the ED value is to 1.

[0093] The voltage level distribution and topology density of the affected equipment entity nodes are retrieved from the risk inference subgraph. Voltage level distribution refers to the statistical count of the number of devices at each voltage level affected by the fault, such as the number of 500kV, 220kV, and 110kV devices affected. Topology density reflects the tightness of connections between affected nodes and can be obtained by calculating the ratio of the actual number of connections between affected nodes to the theoretical maximum number of connections.

[0094] Node group features are constructed based on voltage level distribution and topology connection density. Voltage levels are weighted according to their importance, with higher voltage level devices typically receiving higher weights. For example, a weight of 1.0 can be assigned to 500kV devices, 0.8 to 220kV devices, and 0.6 to 110kV devices. Then, the weighted voltage distribution entropy is calculated, representing the degree of dispersion of the voltage distribution. Simultaneously, combined with topology connection density, a node group feature vector NG = [voltage...]. entropy , connection density ].

[0095] The propagation extension index and node group characteristics are fed into a weight generation function to calculate the propagation weight factor and sensitivity weight factor. The weight generation function uses a soft allocation mechanism to ensure that the sum of the two weight factors is 1. Specifically, the calculation method is: Propagation weight factor WP = sigmoid(α × ED + β × voltage) entropy +γ×connection density The sensitivity weight factor WS = 1 - WP, where α, β, and γ are adjustable hyperparameters used to control the degree of influence of each factor. The sigmoid function ensures that the weight values ​​are limited to a reasonable range.

[0096] The risk propagation intensity feature is weighted by a propagation weighting factor to obtain the weighted risk propagation component. The risk propagation intensity feature RF is a feature vector characterizing the severity of fault propagation, including dimensions such as propagation rate and scope of impact. The weighted risk propagation component is calculated as WRF = WP × RF, which realizes dynamic weight adjustment of the risk propagation feature.

[0097] The weighted security sensitivity component (WSF) is obtained by weighting the security sensitivity feature with the sensitivity weighting factor. The security sensitivity feature SF represents the sensitivity of each part of the system to failure, including dimensions such as load importance and equipment value. The weighted security sensitivity component is calculated as WSF = WS × SF, which realizes dynamic weight adjustment of the security sensitivity feature.

[0098] In the nonlinear fusion module, a nonlinear transformation is applied to the weighted risk propagation component and the weighted security sensitivity component, and the results are accumulated to generate a comprehensive risk feature. The nonlinear transformation adopts a multilayer perceptron structure and introduces a nonlinear activation function to capture the complex interaction between the two types of features. Specifically, the WRF and WSF are first concatenated into a merged feature vector CF=[WRF, WSF]. Then, a nonlinear transformation is performed through a two-layer neural network: H=ReLU(W1×CF+b1), CRF=W2×H+b2, where W1 and W2 are weight matrices, b1 and b2 are bias vectors, and ReLU is a modified linear unit activation function. The final output CRF is the comprehensive risk feature, which is used for subsequent risk assessment and decision support.

[0099] Through the above implementation method, a dynamic weighted nonlinear fusion of risk propagation intensity characteristics and security sensitivity characteristics is realized, so that the fusion weight can be adaptively adjusted according to the hop depth of the fault propagation path and the characteristics of the node group, thereby more accurately assessing the comprehensive risk status of the power system.

[0100] In one optional implementation, based on the risk level assessment results and the current operating status of the power grid, a decision-making instruction set for equipment control commands and topology reconfiguration schemes is output through a reinforcement learning-based decision strategy generation module, including:

[0101] The risk level identifier and the set of affected equipment nodes are read from the risk level assessment results. At the same time, the power flow distribution data and equipment operating parameters are read from the current power grid operating status. Feature encoding is performed on the risk level identifier and the set of affected equipment nodes to obtain a risk state representation. Feature encoding is performed on the power flow distribution data and the equipment operating parameters to obtain an operating state representation. The risk state representation and the operating state representation are concatenated to form a decision state vector. The decision state vector is then sent to the reinforcement learning-based decision strategy generation module.

[0102] The decision state vector is processed by feature extraction and strategy mapping according to the decision strategy generation module, and outputs a device control action sequence and a topology reconstruction action sequence. Based on the device control action sequence, control instructions for the control target device are generated, and based on the topology reconstruction action sequence, a reconstruction scheme for the network topology connection is generated. The control instructions and the reconstruction scheme are combined to form a complete set of autonomous decision instructions.

[0103] The risk level assessment results are retrieved through the data interface of the risk assessment module. Risk level identifiers are represented by integer enumeration values, ranging from 1 to 5, corresponding to five risk levels: extremely low, low, medium, high, and extremely high. The set of affected device nodes is stored using a device identifier list structure, where each device identifier is a unique 32-bit integer code. The set capacity is capped at 500 device nodes to accommodate large-scale power grid scenarios. The data retrieval interface uses a synchronous call method with a timeout of 200 milliseconds. A retry mechanism is triggered upon retrieval failure, with a maximum of 3 retries and a retry interval of 50 milliseconds.

[0104] The current operating status of the power grid is read from two main components: power flow distribution data and equipment operating parameters. Power flow distribution data records the active power, reactive power, voltage amplitude, and phase angle information of each bus node, in floating-point double-precision format. Equipment operating parameters cover information such as generator output, transformer tap position, circuit breaker opening and closing status, and load power. Parameters are acquired at a frequency of 10 times per second, with a data buffer window of 300 seconds. The acquisition of power flow distribution data and equipment operating parameters is achieved through a data acquisition and monitoring interface. The interface supports transmission network communication protocols, and data transmission uses a transmission control protocol connection method to ensure reliability.

[0105] The risk level identifier and the feature encoding of the affected device node set are implemented using a combination of one-hot encoding and embedding vectors. The risk level identifier is encoded using a 5-dimensional one-hot vector, and the affected device node set is formed by concatenating the device type embedding vector and the location encoding vector to form a 128-dimensional feature vector. The device type embedding vector has a dimension of 64 and is obtained through a pre-trained device feature model. The location encoding vector also has a dimension of 64 and is calculated based on the connectivity and centrality indices of the device in the power grid topology. The issue of the variable number of affected device nodes is addressed through an attention mechanism, which aggregates the feature vectors of all device nodes by weighted summation to form a fixed-dimensional set representation vector.

[0106] The feature encoding of power flow distribution data and equipment operating parameters is implemented using a multilayer perceptron network. The input layer dimension is dynamically determined according to the power grid scale, the hidden layer uses a modified linear unit activation function, and the output layer dimension is fixed at 256 dimensions. The power flow distribution data includes a feature dimension of 4 times the number of buses, corresponding to four parameters: active power, reactive power, voltage amplitude, and phase angle. Data normalization uses the maximum-minimum standardization method. Equipment operating parameters include status information of four types of equipment: generators, transformers, circuit breakers, and loads. The parameter dimension of each type of equipment is determined according to the number of equipment and parameter type. Normalization uses the standard fraction standardization method.

[0107] The concatenation operation between the risk state representation and the operational state representation involves chaining the two vectors together along their feature dimensions. The risk state representation has 133 dimensions, the operational state representation has 256 dimensions, and the concatenated decision state vector has 389 dimensions. The concatenation operation is implemented using array concatenation functions from a numerical computation library, with the concatenation axis set to the last dimension to ensure dimensionality consistency during batch processing. Before being fed into the reinforcement learning module, the decision state vector undergoes data type conversion and device migration, converting the numerical array to tensor format and setting the data type to 32-bit floating-point to balance precision and performance requirements.

[0108] The reinforcement learning-based decision policy generation module is implemented using a deep Q-network architecture. The network structure includes an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the hidden layers are 512, 256, and 128, respectively, and the activation function is a rectified linear unit (RCU). The output layer uses a linear activation function. Network training employs an experience replay mechanism, with an experience pool capacity of 100,000 experience samples, a batch size of 64, a learning rate of 0.001, and a discount factor of 0.95. The target network is updated softly every 1000 steps, with an update coefficient of 0.005.

[0109] Feature extraction of the decision state vector is achieved through forward propagation computation. The input 389-dimensional vector undergoes a nonlinear transformation through three hidden layers to extract high-level semantic feature representations. Batch normalization is used to stabilize the training process during feature extraction, and the random deactivation ratio is set to 0.2 to prevent overfitting. Policy mapping processes map the extracted features to equipment control action sequences and topology reconstruction action sequences. The number of neurons in the output layer is equal to the dimension of the action space. The equipment control action space includes operation types such as generator output adjustment, transformer tap adjustment, and capacitor switching, while the topology reconstruction action space includes topology change operations such as circuit breaker opening and closing and tie switch operation.

[0110] The generation of equipment control action sequences employs a continuous action space discretization method. Generator output adjustment actions are divided into five discrete actions: increase by 10%, increase by 5%, remain unchanged, decrease by 5%, and decrease by 10%. Transformer tap changer adjustment actions are divided into three discrete actions: raise by one level, remain unchanged, and lower by one level. The action sequences are represented using multi-dimensional action vectors, where the vector dimension equals the number of controllable devices. Each dimension corresponds to the control action code for one device. Action codes use integer enumeration values, with the value range determined based on the device type and control range. A code value of 0 indicates maintaining the status quo, a positive value indicates an increase or raise operation, and a negative value indicates a decrease or lower operation.

[0111] The generation of topology reconfiguration action sequences is based on the operational space definition of circuit breakers and switching devices. Each circuit breaker or switch corresponds to a binary action, where 0 represents an opening operation and 1 represents a closing operation. The topology reconfiguration action sequence is represented by a binary vector, with the vector dimension equal to the total number of operable circuit breakers and switches. Each element in the vector corresponds to the operational state of a device. The constraints of topology reconfiguration include radial constraints, connectivity constraints, and device capacity constraints. During the action sequence generation process, a constraint checking module ensures that the generated actions meet the safety requirements of power grid operation.

[0112] The generation of control commands is based on a mapping table between device control action sequences and device identifiers. This mapping table records the correspondence between action codes and specific control commands. Control commands adopt a standardized command format, including fields such as target device identifier, operation type, target value, execution time, and priority. The target device identifier uses a unique device code, the operation type uses a string enumeration value, the target value uses floating-point data to represent the desired device parameter value, the execution time uses a timestamp format to specify the moment the command is executed, and the priority uses an integer value to represent the urgency of the command execution.

[0113] The reconfiguration scheme is generated based on the topology reconfiguration action sequence and network topology information. The scheme includes a switching operation sequence, operation timing arrangement, and safety check results. The switching operation sequence records the circuit breaker and switching operations to be performed, with each operation including equipment identification, operation type, operation time, and operation conditions. The operation timing arrangement determines the execution order and time interval of each operation according to power grid operation safety requirements, avoiding power grid instability caused by simultaneous operations. The safety check results record the safety assessment of the reconfiguration scheme, including indicators such as power flow feasibility, voltage stability, and frequency stability.

[0114] The combination of control commands and reconfiguration schemes is implemented using an instruction set data structure. The instruction set includes a control command list, reconfiguration scheme objects, execution sequence planning, and rollback strategies. The control command list records all necessary device control operations, the reconfiguration scheme objects record detailed information about network topology changes, the execution sequence planning determines the coordinated execution order of control commands and reconfiguration operations, and the rollback strategy defines recovery measures in abnormal situations. The autonomous decision-making instruction set uses a structured data format for serialized storage and transmission, ensuring cross-platform compatibility and readability.

[0115] In one optional implementation, the decision strategy generation module maps historical decision execution results to the weight parameter space of the time-series causal reasoning rule for adjustment, executes the autonomous decision instruction set, and updates the power grid operation status data after decision execution to the corresponding node attributes of the knowledge graph, including:

[0116] The feature quantities of historical decision execution results are mapped to the weight parameter space of the time-series causal reasoning rule, and the weight parameters are adjusted accordingly based on the mapping results; based on the adjusted weight parameters, an autonomous decision instruction set including equipment control instructions and topology reconfiguration schemes is executed; the power grid operation status data after the decision instructions are executed is obtained, and the power grid operation status data is updated and written into the attribute fields of the corresponding nodes in the knowledge graph.

[0117] The feature quantities of historical decision execution results are mapped to the weight parameter space of time-series causal inference rules. Specifically, decision execution data over a past period (e.g., the last 30 days) is extracted from the knowledge graph, including equipment control commands, topology reconfiguration schemes, and their corresponding execution effects. This historical data contains multi-dimensional features, such as the rate of change of grid load, changes in equipment operating status, and the degree of improvement in system stability indicators. For each historical decision record, its feature vector F is extracted, which contains dimensional information such as decision type, grid state at the time of execution, and changes in grid state after execution.

[0118] For time-series causal inference rules, the weight parameter space is represented by W, containing the importance weights of different decision rules. A mapping function M(F)→W is constructed to map the feature vector F to the weight adjustment amount ΔW. During the mapping process, a gradient-based learning method is used to calculate the difference between the historical decision effect score S and the expected effect S', and the weight parameters are adjusted by minimizing this difference. Specifically, an adaptive weighting method is used: for decision types that produce positive effects, the weight of the corresponding rule is increased; for decision types with poor effects, the weight of the corresponding rule is decreased.

[0119] For example, when historical data shows that adjusting the reactive power compensation device of a substation can effectively improve voltage stability under specific load conditions, the weights of the related causal rules will be enhanced. After the mapping process is completed, the weight parameters of the time-series causal inference rules are updated, forming a new weight set W'.

[0120] Based on the adjusted weight parameters, an autonomous decision-making instruction set is executed to obtain current power grid operating status data from the knowledge graph, including information such as node voltage and line power flow. Then, based on the updated temporal causal reasoning rule weights W', the current power grid state is evaluated and a decision candidate set D is generated. This candidate set contains various equipment control instructions and topology reconfiguration schemes, such as transformer tap adjustment, capacitor bank switching, and line opening and closing operations.

[0121] For each decision di in the candidate set, a time-series causal model is used to predict the power grid state after its execution, and a comprehensive score Ei is calculated based on multi-dimensional indicators such as safety, economy, and reliability. The calculation formula can be simplified to Ei = Σ(Wj' × Pij), where Wj' is the updated rule weight, and Pij is the expected performance of decision di on indicator j. Based on the comprehensive score, the decision scheme with the highest score is selected from the candidate set D to form the final autonomous decision instruction set D×.

[0122] During the execution of the decision instruction set D×, the system monitors the execution status and records the actual effect of each operation. For example, after executing a bus segmentation operation, it records data such as voltage changes on both sides of the bus and power flow redistribution in real time. If an anomaly is detected during execution, such as equipment response timeout or grid parameters exceeding safe ranges, an emergency decision-making mechanism will be activated to adjust or terminate the execution of subsequent instructions.

[0123] The system acquires power grid operation status data after the execution of decision-making instructions and updates it to the knowledge graph. After the decision is executed, the system collects the latest power grid operation status data, including key indicators such as node voltage values, line power flow, equipment operating status, and system frequency. This data is compared and analyzed with the state before execution to generate a decision execution effect report.

[0124] For knowledge graph updates, the nodes requiring updates are located, such as the operated device nodes, related line nodes, and their associated nodes. Then, a data update package is constructed, containing node identifiers, attribute field names, new values, timestamps, and other information. The update package is written to the corresponding node's attribute fields via the knowledge graph interface. For example, for an operation to adjust the tap of a transformer, its "current tap position," "last adjustment time," and "adjustment reason" attribute fields in the knowledge graph are updated. Simultaneously, attributes such as "voltage level" and "power flow" of related line and bus nodes are also updated.

[0125] After the update, the power grid status information in the knowledge graph remains up-to-date, providing accurate foundational data for the next round of decision-making. The entire process forms a closed loop: historical decision results influence future decisions through weight parameter adjustments, and the results of decision execution update the knowledge graph as new historical data, thus achieving continuous optimization of the power system's autonomous decision-making capabilities.

[0126] Through the above implementation methods, the power system can continuously optimize decision-making strategies based on historical experience, improve the accuracy and effectiveness of autonomous decision-making, and maintain the timeliness and integrity of knowledge graph data, thus providing strong support for the safe and stable operation of the power grid.

[0127] A second aspect of the present invention provides a power grid operation risk intelligent identification system, comprising:

[0128] The first module is used to acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph, identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges.

[0129] The second module is used to assess the risk level based on the risk source nodes and impact range information in the risk reasoning subgraph, combined with the power grid operation constraints and safety margin requirements, and obtain the risk level assessment result.

[0130] The third module is used to output an autonomous decision-making instruction set of equipment control instructions and topology reconfiguration schemes based on the risk level assessment results and the current operating status of the power grid through a reinforcement learning-based decision strategy generation module. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the autonomous decision-making instruction set, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

[0131] A third aspect of the present invention provides an electronic device, comprising:

[0132] processor;

[0133] Memory used to store processor-executable instructions;

[0134] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0135] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0136] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent identification of power grid operation risks, characterized in that, include: Acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph to identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges; Based on the risk source nodes and impact range information in the risk reasoning subgraph, and combined with power grid operation constraints and safety margin requirements, the risk level is assessed to obtain the risk level assessment results, including: The device importance of the risk source node and the topology depth of the fault propagation path are extracted from the risk reasoning subgraph. The device importance and the topology depth are adaptively coupled and calculated, and an electrical distance attenuation factor is introduced to form a risk propagation intensity feature that includes spatial attenuation characteristics. The affected equipment entity nodes in the scope of impact information and the operating nodes in the power grid operation constraints are identified through multi-dimensional intersection. For the node group in the intersection, the dynamic deviation between its current operating status and the hierarchical safety margin requirements is calculated. At the same time, the coupling correlation strength between nodes is considered to construct a safety sensitivity feature with group characteristics. The risk propagation intensity feature and the security sensitivity feature are input into the nonlinear fusion module. During the fusion process, the fusion weight is dynamically adjusted according to the hop depth of the fault propagation path and the node group feature in the risk reasoning subgraph to generate a risk level assessment result that reflects the overall risk situation of the system. Based on the risk level assessment results and the current operating status of the power grid, the decision strategy generation module based on reinforcement learning outputs a set of autonomous decision-making instructions for equipment control and topology reconfiguration schemes. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the set of autonomous decision-making instructions, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

2. The method according to claim 1, characterized in that, Multi-hop reasoning based on temporal causal reasoning rules is performed on the entity nodes and relational edges in the knowledge graph to identify fault propagation paths and risk association chains, generating a risk reasoning subgraph containing risk source nodes, propagation paths, and impact scope, including: In the knowledge graph, device entity nodes whose state attributes have changed abnormally are detected, and the detected device entity nodes, their timestamps, and state change features are encapsulated to form initial inference anchors. Starting from the initial inference anchor point, the system traverses outward along the association edges in the knowledge graph. It then verifies the causal transitivity and temporal sequence of the subsequent device entity nodes connected by the association edges using temporal causal inference rules. Finally, it combines the verified subsequent device entity nodes and their association edges with the preceding nodes to form a propagation path segment. The subsequent device entity nodes in the propagation path segment are updated to new inference anchor points. The aforementioned traversal and verification process is repeated. Propagation path segments are continuously superimposed until the subsequent device entity nodes connected by the association edge cannot pass the verification of the temporal causal inference rule. The superimposed propagation path segments are connected to form multiple complete fault propagation paths starting from the initial inference anchor point. Based on the device entity nodes and associated edges contained in the multiple complete fault propagation paths, a risk reasoning subgraph is constructed.

3. The method according to claim 2, characterized in that, The causal transitivity and temporal sequence of the subsequent device entity nodes connected by the associated edges are verified using temporal causal reasoning rules. The verified subsequent device entity nodes and their associated edges with the preceding nodes are combined to form a propagation path segment, including: The state change timestamps and abnormal state features of the preceding nodes are obtained from the inference anchor point, and the state change timestamps and abnormal state features of the connected subsequent device entity nodes are obtained through the association edges based on the preceding nodes. The time interval is calculated based on the state change timestamps of the preceding node and the following device entity node. The time interval is compared with the physical propagation delay of the associated edge to complete the temporal sequence verification. At the same time, the abnormal state characteristics of the preceding node and the following device entity node are input into a preset causal transit rule base for matching to complete the causal transitivity verification. For node pairs that have completed both temporal sequence verification and causal transitivity verification, the subsequent device entity node is bound to its corresponding association edge. During the binding process, the time interval value and the causal rule matching result are written into the attribute field of the association edge, ultimately forming a propagation path segment containing complete verification information.

4. The method according to claim 1, characterized in that, The risk propagation intensity feature and the security sensitivity feature are input into the nonlinear fusion module. During the fusion process, the fusion weights are dynamically adjusted based on the hop depth of the fault propagation path and the node group features in the risk reasoning subgraph, including: The risk reasoning subgraph traverses all fault propagation paths and obtains the hop depth of each path. Based on the hop depth, the propagation extension index is calculated. At the same time, the voltage level distribution and topology connection density of the affected equipment entity nodes are read from the risk reasoning subgraph. Based on the voltage level distribution and topology connection density, node group features are constructed. The propagation extension index and the node group features are fed into the weight generation function and the propagation weight factor and sensitivity weight factor are calculated. The risk propagation intensity feature is weighted by the propagation weight factor to obtain a weighted risk propagation component, and the security sensitivity feature is weighted by the sensitivity weight factor to obtain a weighted security sensitivity component. In the nonlinear fusion module, a nonlinear transformation is applied to the weighted risk propagation component and the weighted security sensitivity component, and they are accumulated to generate a comprehensive risk feature.

5. The method according to claim 1, characterized in that, Based on the risk level assessment results and the current operating status of the power grid, the decision-making instruction set for equipment control commands and topology reconfiguration schemes, generated by the reinforcement learning-based decision-making strategy generation module, includes: The risk level identifier and the set of affected equipment nodes are read from the risk level assessment results. At the same time, the power flow distribution data and equipment operating parameters are read from the current power grid operating status. Feature encoding is performed on the risk level identifier and the set of affected equipment nodes to obtain a risk state representation. Feature encoding is performed on the power flow distribution data and the equipment operating parameters to obtain an operating state representation. The risk state representation and the operating state representation are concatenated to form a decision state vector. The decision state vector is then sent to the reinforcement learning-based decision strategy generation module. The decision state vector is processed by feature extraction and strategy mapping according to the decision strategy generation module, and outputs a device control action sequence and a topology reconstruction action sequence. Based on the device control action sequence, control instructions for the control target device are generated, and based on the topology reconstruction action sequence, a reconstruction scheme for the network topology connection is generated. The control instructions and the reconstruction scheme are combined to form a complete set of autonomous decision instructions.

6. The method according to claim 1, characterized in that, The decision strategy generation module maps historical decision execution results to the weight parameter space of the time-series causal reasoning rule for adjustment, executes the autonomous decision instruction set, and updates the power grid operation status data after decision execution to the corresponding node attributes of the knowledge graph, including: The feature quantities of historical decision execution results are mapped to the weight parameter space of the time-series causal reasoning rule, and the weight parameters are adjusted accordingly based on the mapping results; based on the adjusted weight parameters, an autonomous decision instruction set including equipment control instructions and topology reconfiguration schemes is executed; the power grid operation status data after the decision instructions are executed is obtained, and the power grid operation status data is updated and written into the attribute fields of the corresponding nodes in the knowledge graph.

7. A power grid operation risk intelligent identification system, used to implement the method of any one of claims 1-6, characterized in that, include: The first module is used to acquire power grid operation status data and topology data, and construct a knowledge graph based on a predefined ontology model; perform multi-hop reasoning based on temporal causal reasoning rules on entity nodes and related edges in the knowledge graph, identify fault propagation paths and risk association chains, and generate a risk reasoning subgraph containing risk source nodes, propagation paths and impact ranges. The second module is used to assess the risk level based on the risk source nodes and impact range information in the risk reasoning subgraph, combined with the power grid operation constraints and safety margin requirements, and obtain the risk level assessment result. The third module is used to output an autonomous decision-making instruction set of equipment control instructions and topology reconfiguration schemes based on the risk level assessment results and the current operating status of the power grid through a reinforcement learning-based decision strategy generation module. The decision strategy generation module maps the historical decision execution results to the weight parameter space of the time-series causal reasoning rules for adjustment, executes the autonomous decision-making instruction set, and updates the power grid operating status data after the decision execution to the corresponding node attributes of the knowledge graph.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images