A power grid fault fast recovery system and method based on graph data detection

Abstract

This invention discloses a rapid power grid fault recovery system and method based on graph data detection, belonging to the field of power system automation. The system includes: a topology analysis module, which constructs a weighted graph of the power grid and calculates spectral features to generate a topology metric for quantifying grid resilience; a data consistency detection module, which calculates the geometric consistency between local measurements and global state data to generate a consistency score; a neural symbolic decision-making module, which integrates the topology metric and the consistency score, generates high-level control intentions through reinforcement learning, and verifies and corrects them using a symbolic rule base that encodes physical constraints; and a fault recovery execution module, which generates equipment control commands to achieve phased fault isolation and power restoration. This invention improves the real-time performance, data reliability, and decision security of power grid fault response, possesses autonomous learning and adaptive decision-making capabilities, and can effectively enhance the self-healing capability and operational resilience of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power system automation and intelligent control technology, and more specifically to a system and method for rapid recovery of power grid faults based on graph data detection. Background Technology

[0002] The self-healing capability of a power grid is a core feature of modern smart grids, referring to the system's ability to automatically detect anomalies, quickly isolate faults, reconstruct a healthy topology, and restore power supply after suffering a fault or disturbance. With the large-scale integration of new energy sources such as wind power and photovoltaics, the operating conditions of power grids are becoming increasingly complex, and the types of faults are becoming more diverse. Traditional protection and control systems based on preset rules are unable to cope with complex fault scenarios, and there is an urgent need for intelligent self-healing systems with autonomous learning and adaptive decision-making capabilities.

[0003] Existing power grid self-healing systems suffer from numerous technical bottlenecks. First, their topology analysis capabilities are insufficient. Traditional methods primarily rely on static N-1 safety checks and offline simulation analysis, lacking effective means for real-time quantitative analysis of the power grid topology, making it difficult to meet millisecond- to second-level online response requirements. Second, their measurement data reliability verification capabilities are inadequate. Massive amounts of data from SCADA, PMU, and other devices are difficult to process efficiently. Traditional bad data detection methods show decreased accuracy when multiple anomalies or topology misjudgments occur, lacking an effective verification mechanism for the consistency between local measurement data and global state data. Third, the embedding of physical constraints in intelligent control decisions is insufficient. Control strategies based on reinforcement learning and other artificial intelligence methods exhibit "black box" characteristics, poor interpretability, and difficulty in effectively integrating physical constraints into the decision-making process, leading to compromised decision security. Finally, the fault recovery process lacks systematic management. Traditional methods lack a clear phased coordination mechanism for fault isolation and power restoration, resulting in poor controllability and predictability of the recovery process.

[0004] Therefore, there is an urgent need for a power grid self-healing control system and method that can quantify the topology in real time, effectively verify the consistency of measurement data, embed physical constraints, and have phased recovery management capabilities. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a grid fault rapid recovery system and method based on graph data detection that overcomes or at least partially solves the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a power grid fault rapid recovery system based on graph data detection, comprising: The topology analysis module is used to acquire the topology data of the power grid, construct a weighted graph characterizing the topology of the power grid, and calculate the spectral features of the weighted graph to generate a topology metric for quantifying the resilience of the power grid. The data consistency detection module is used to acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and to calculate the geometric consistency between each local measurement data and the global state data of the corresponding area to generate a consistency score to indicate data quality. The neural symbolic decision module is used to receive the topology metric and the consistency score, fuse them based on a pre-trained reinforcement learning model to generate a high-level control intention, and use a symbolic rule library encoded with physical constraints to verify and correct the high-level control intention to generate control actions. The fault recovery execution module is used to generate and issue specific equipment control commands based on the control actions, so as to coordinate the operation of circuit breakers, topology reconfiguration and load balancing in the power grid, and realize phased fault isolation and power supply restoration.

[0007] Preferably, the topology analysis module is further used to calculate the eigenvalues ​​of the Laplacian matrix of the weighted graph, and to use the second smallest eigenvalue as an algebraic connectivity index to quantify the risk of power grid splitting.

[0008] Preferably, the data consistency detection module is further configured to map the local measurement data and the global state data into vectors, and obtain the consistency score by calculating the normalized inner product between the vectors.

[0009] Preferably, the neural symbolic decision-making module further includes: The RL agent submodule is used to output the higher-level control intent based on the current system state; A symbol rule library is used to store if-then rules that encode at least one physical constraint among generator ramp rate, circuit breaker operating time, and transformer tap change interval. The rules engine is used to trigger the correction of the higher-level control intent or trigger the system to enter a degraded operation mode when the higher-level control intent violates the rules in the symbol rule base.

[0010] Preferably, the fault recovery execution module is specifically used to execute the following phased process: During the fault detection phase, faults are identified and located based on abrupt changes in the topology metric and / or anomalies in the consistency score. During the fault isolation phase, based on the topology metric and the consistency score, the isolation priority of the faulty line is determined, and the circuit breaker is coordinated to perform isolation actions. During the topology reconfiguration phase, optimization algorithms are used to search for reconfiguration paths and optimize power allocation on the remaining healthy lines. During the load recovery phase, the disconnected load is gradually restored, and a safety check is performed during the recovery process.

[0011] Secondly, embodiments of the present invention provide a method for rapid recovery of power grid faults based on graph data detection, applicable to any of the above-mentioned systems, comprising the following steps: S1. Obtain the topology data of the power grid, construct a weighted graph representing the topology of the power grid, and calculate the spectral features of the weighted graph to generate a topology metric for quantifying the resilience of the power grid. S2. Acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and calculate the geometric consistency between each of the local measurement data and the global state data of the corresponding region to generate a consistency score for indicating data quality. S3. Based on the pre-trained reinforcement learning model, the topology metric and the consistency score are fused to generate a high-level control intention, and the high-level control intention is verified and corrected using a symbol rule base encoded with physical constraints to generate control actions. S4. Based on the control actions, generate and issue specific equipment control commands to coordinate circuit breaker actions, topology reconfiguration, and load balancing in the power grid, thereby achieving phased fault isolation and power supply restoration.

[0012] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: (1) Real-time quantization capability of topology This invention acquires power grid topology data through a topology analysis module, constructs a weighted graph, calculates spectral features, and generates a topology metric for quantifying power grid resilience. This metric can reflect the structural characteristics of the power grid in real time, providing a quantitative basis for fault detection and recovery decisions. Compared with traditional offline simulation analysis, the response speed is significantly improved, meeting the timeliness requirements of real-time control systems. (2) Measurement data consistency verification capability This invention acquires local measurement data and global state data through a data consistency detection module, calculates the geometric consistency between the two, and generates a consistency score to indicate data quality. This method can effectively identify abnormal measurement data and topology misjudgment scenarios, providing a reliable data foundation for subsequent control decisions. (3) The ability to integrate physical constraints in intelligent decision-making This invention receives topological metrics and consistency scores through a neural symbolic decision-making module, fuses them using a reinforcement learning model to generate a high-level control intent, and then verifies and corrects the high-level control intent using a symbolic rule base encoded with physical constraints to generate control actions. This approach maintains the flexibility of AI decision-making while ensuring that control actions conform to physical constraints, thus improving the safety and interpretability of the decision-making process. (4) Phased fault recovery management capability This invention utilizes a fault recovery execution module to generate and issue equipment control commands based on control actions, coordinating circuit breaker actions, topology reconfiguration, and load balancing to achieve phased fault isolation and power restoration. This phased mechanism clarifies the timing and functional boundaries of the recovery process, improving the controllability and predictability of the fault recovery process. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the overall system architecture provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the data consistency detection module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the neural symbolic decision-making module architecture provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall process of the method provided in the embodiments of the present invention. Detailed Implementation

[0015] 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.

[0016] like Figure 1 As shown, this embodiment of the invention discloses a rapid power grid fault recovery system based on graph data detection, comprising: The topology analysis module is used to acquire the topology data of the power grid, construct a weighted graph characterizing the topology of the power grid, and calculate the spectral features of the weighted graph to generate topology metrics for quantifying the resilience of the power grid. The data consistency detection module is used to acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and to calculate the geometric consistency between each local measurement data and the global state data of the corresponding area to generate a consistency score to indicate data quality. The neural symbolic decision module receives topological metrics and consistency scores, fuses them based on a pre-trained reinforcement learning model to generate high-level control intentions, and uses a symbolic rule base encoded with physical constraints to verify and correct the high-level control intentions, generating control actions. The fault recovery execution module is used to generate and issue specific equipment control commands based on control actions, so as to coordinate the operation of circuit breakers, topology reconfiguration and load balancing in the power grid, and realize phased fault isolation and power supply restoration.

[0017] In one embodiment, the topology analysis module is further used to calculate the eigenvalues ​​of the Laplacian matrix of the weighted graph, and the second smallest eigenvalue is used as an algebraic connectivity index to quantify the risk of grid splitting.

[0018] Specifically, the topology analysis module is used to calculate in real time indicators such as algebraic connectivity, vulnerability index, edge criticality, and bridge-side detection of the power grid topology. The topology analysis module receives topology data from the physical power grid layer (including parameters such as bus sets, line sets, and line reactance) and outputs topology metrics (including...). , , The topology analysis module (including the bridge edge list) is passed to the fault detection module for mutation detection and to the fault isolation module for calculating isolation priorities. When topology metrics exceed preset thresholds, an alarm is triggered and the fault recovery process is initiated. The topology analysis module includes a graph construction submodule, a Laplace matrix calculation submodule, and an algebraic connectivity calculation submodule.

[0019] (1) Graph construction submodule: The graph construction submodule constructs a weighted graph from the power grid topology data. ,in For the set of busbars, This is a set of routes. For each route... Edge weight is defined as ,in Line reactance (unit: per unit, range: 0.01 to 0.5). Edge weights reflect the electrical distance of the line; the smaller the reactance, the larger the edge weight, indicating a tighter electrical connection. The system maintains an adjacency matrix. ,in If a line exists ,otherwise Degree matrix It is a diagonal matrix. Represents a node The weighting degree.

[0020] (2) Laplace matrix calculation submodule: The Laplacian matrix calculation submodule constructs the normalized Laplacian matrix. The matrix has the following properties: Let be a symmetric positive semi-definite matrix, and let its eigenvalues ​​satisfy... , Corresponding to a feature vector consisting entirely of 1s The system computes the graph if and only if the graph is connected. The Lanczos iterative algorithm is used. The former The smallest eigenvalues ​​(typical values) The computational complexity is ,in Let be the number of edges. For the IEEE 118-node system ( , The calculation time is approximately 15 milliseconds.

[0021] (3) Algebraic connectivity calculation submodule: The algebraic connectivity computation submodule extracts the second smallest eigenvalue of the Laplacian matrix. (Fiedler value). The connectivity strength of the graph was quantified. A larger value indicates that the network is less likely to split into isolated islands. For power grid topologies, The physical meaning is: when removing certain lines causes the graph to split, It can suddenly drop to near 0. (System setting threshold) ,when An islanding alarm is triggered at any time. Experiments show that in an N-1 fault scenario, The average decrease was 10% to 30% in the N-2 fault scenario. The average decline is 30% to 60% when a critical path failure leads to islanding. It dropped to less than 0.01.

[0022] To more comprehensively assess the overall vulnerability of the power grid, in a preferred embodiment of the present invention, the topology analysis module calculates the algebraic connectivity. Building upon this foundation, a vulnerability index calculation submodule is also included. This submodule comprehensively considers algebraic connectivity and bridge-edge ratio to define the vulnerability index. ,in This is a regularization constant (to prevent division by zero). The bridge edge ratio (this bridge edge ratio is calculated from the list of bridge edges detected by the bridge edge detection submodule below, specifically expressed as the number of bridge edges / the total number of edges, with a value range of 0 to 1). A higher value indicates a more fragile network. Bridge edges are edges that, when removed, cause the graph to become disconnected; their existence indicates a single point of failure risk in the network. The system sets a vulnerability threshold. ,when If a vulnerability alert is triggered, it is recommended to strengthen monitoring or perform preventative maintenance.

[0023] To identify the critical lines in the power grid that have the greatest impact on the overall network resilience, in another preferred embodiment of the invention, the topology analysis module further includes an edge criticality assessment submodule, which calculates the impact of a single line on the overall network resilience. For each line... Define edge criticality ,in The algebraic connectivity of the original graph. To remove edges Algebraic connectivity afterward. A higher value indicates a more critical path, and removing it will have a greater impact on connectivity (value range: 0 to 1). The system calculates the edge criticality of all paths, sorts them in descending order, and identifies the top 10% of critical paths as key monitoring targets. The computational complexity is O(log n). For the IEEE 118-node system, the computation time is approximately 30 milliseconds.

[0024] To identify single-point-of-failure risks in the power grid, in another preferred embodiment of the invention, the topology analysis module further includes a bridge-edge detection submodule. This submodule employs a depth-first search (DFS) algorithm to identify bridge edges. The algorithm maintains the access timestamp and earliest reachable timestamp for each node. When an edge exists... Make The earliest reachable timestamp is greater than When accessing the timestamp, It is located on the side of the bridge. The algorithm's time complexity is O(n log n). For an IEEE 118-node system, the computation time is approximately 5 milliseconds. The detected bridge edge list is output to the vulnerability index calculation submodule and the edge criticality assessment submodule for a comprehensive evaluation of network resilience.

[0025] It is understood that the above preferred embodiments can be combined arbitrarily. For example, a more comprehensive topology analysis module can simultaneously include an algebraic connectivity calculation submodule, a vulnerability index calculation submodule, an edge criticality assessment submodule, and a bridge edge detection submodule, whose output topology metrics (including...) , , (And the bridgeside list) can be used simultaneously for mutation detection, isolation priority calculation and preventive maintenance decision-making, thereby achieving a multi-dimensional quantitative assessment of power grid resilience from the overall to the local.

[0026] In one embodiment, the data consistency detection module is further configured to map local measurement data and global state data into vectors, and obtain a consistency score by calculating the normalized inner product between the vectors.

[0027] Combination Figure 2 As shown, specifically, the data consistency detection module is used to model local measurement data into local data fragments, and to realize bad data detection and topology misjudgment identification by calculating the consistency score between local data and global state estimation. The data consistency detection module is used to calculate the consistency score for each measurement point. This module first includes a local data construction submodule and a global data projection submodule to prepare data for subsequent calculations. (1) Local data construction: The local data construction submodule generates local patches from local measurement data. For each bus... Define a local patch For including busbars Sub-maps of the busbars and their adjacent busbars. Measurement data. (Voltage amplitude, phase angle, active power, reactive power) constitute local data ,in For patch The system supports SCADA low-frequency measurements (4 to 10-second cycles) and PMU high-frequency measurements (50 to 60 frames per second), achieving multi-rate data fusion through timestamp alignment and linear interpolation.

[0028] (2) Global data projection: The global data projection submodule extracts the global state data of the corresponding bus from the state estimation results. State estimation employs either Weighted Least Squares (WLS) or Extended Kalman Filtering, outputting a global state vector. For busbars Local patch Global data State vector exist The projection on The projection operation is implemented through index mapping, with a computational complexity of O(n log n). .

[0029] (3) Calculation of consistency score: The consistency score calculation submodule calculates local data. With global data Consistency. The consistency score is defined as the normalized inner product:

[0030] in For inner product, It is an L2 norm. The range of the consistency score is... A score close to 1 indicates a high degree of consistency between the measurement and estimation; a score close to 0 indicates orthogonality (no correlation); and a score less than 0 indicates inverse correlation (potentially bad data or topological misjudgment). The system sets a consistency threshold. ,when Anomaly detection is triggered at any time.

[0031] To automate the processing of low-quality data and prevent it from affecting control decisions, in a preferred embodiment of the present invention, the data consistency detection module further includes an anomaly detection and weight adjustment submodule. The anomaly detection submodule analyzes buses with consistency scores below a threshold. The system maintains a consistency history record; when the consistency score of a bus is below a certain threshold for three consecutive periods... When this occurs, it is marked as potentially bad data or a topology misjudgment. The weight adjustment submodule dynamically adjusts the measurement weights: for consistency scores of... The measurement weights were adjusted to ,in The original weights are set to 0.1 (based on measurement accuracy), with 0.1 as the minimum weight lower bound (to prevent complete exclusion). The adjusted weights are used for the WLS solution of the next state estimation, reducing the impact of low-consistency measurements on the global estimation. When the consistency score is less than 0.2, the system blocks control actions based on that measurement and triggers a manual review process to ensure that the system does not malfunction due to erroneous data.

[0032] In one embodiment, the neural symbol decision-making module further includes: The RL agent submodule is used to output high-level control intents based on the current system state. A symbol rule library is used to store if-then rules that encode at least one physical constraint among generator ramp rate, circuit breaker operating time, and transformer tap change interval. The rules engine is used to trigger corrections to higher-level control intentions or to trigger the system to enter a degraded operating mode when a higher-level control intention violates a rule in the symbol rule base.

[0033] Specifically, combined Figure 3 As shown, the neural symbolic decision-making module is used to fuse high-level control intentions from reinforcement learning with physical constraints from symbolic rules to generate intelligent and safe control actions. The neural symbolic decision-making module receives topology metrics from the topology analysis module and consistency scores from the data consistency detection module, as well as the current system state vector. (Including 16 dimensions of indicators such as average utilization, maximum utilization, congested line ratio, faulty line ratio, voltage over-limit ratio, and frequency deviation), output control actions. The interpretable reasoning link is passed to the fault recovery execution module for circuit breaker action and load balancing. When an RL action violates physical constraints, constraint correction or degradation mode is triggered. This neural symbolic decision module includes an RL agent submodule, a symbolic rule base, a rule engine, a constraint correction submodule, and a degradation mode trigger.

[0034] The RL agent submodule supports three algorithms: PPO (Proximal Policy Optimization), DQN (Deep Q-Network), and SimpleQ (Table-based Q-learning). PPO employs a policy gradient approach, ensuring training stability by limiting the policy update magnitude, making it suitable for continuous control scenarios. DQN uses a value function approach, improving sample efficiency through empirical replay and a target network, making it suitable for discrete action spaces. SimpleQ uses table-based Q-learning, requiring no deep learning framework and exhibiting inference latency of less than 1 millisecond, making it suitable for scenarios with limited edge resources.

[0035] The state space of the RL agent is a 16-dimensional vector, including indicators such as average utilization, maximum utilization, congested line ratio, faulty line ratio, voltage over-limit ratio, frequency deviation, reserve margin, and renewable energy ratio. The action space includes six discrete actions: Monitor (observe without intervention), Reduce load (reduce congested line traffic by 20%), Reroute (redistribute power), Emergency shed (emergency load shedding by 40%), Increase reserve (increase spinning reserve), and Voltage regulate. The reward function comprehensively considers congestion mitigation, utilization improvement, stability rewards, unnecessary action penalties, and fault penalties, with weights of 2.0, 1.0, 0.5, -0.3, and -8.0, respectively.

[0036] To ensure that the actions output by the RL comply with the operational constraints of the power grid's physical equipment, the neural symbolic decision-making module also includes a symbolic rule base and a rule engine. The symbolic rule base encodes physical constraints as if-then rules, with the rule format being IF (conditional expression) THEN (action correction) PRIORITY (priority). Typical rules cover physical constraints such as generator ramp rate limits, circuit breaker cooling time constraints, and transformer tap changer intervals. Mandatory rules include that generator output changes must not exceed 3% of rated power per minute (otherwise, ramp rate is limited to 3%), the last circuit breaker action time must be greater than 100 milliseconds (otherwise, action is blocked), and the last transformer tap changer adjustment time must be greater than 30 seconds (otherwise, adjustment is blocked). High-priority rules include prioritizing voltage regulation actions when voltage exceeds the limit by more than 0.1 pu, and prioritizing frequency regulation actions when frequency deviation exceeds 0.2 Hz. The rule engine uses forward chain reasoning, evaluating rules in priority order. For the RL output action a_RL, the rule engine checks whether it violates mandatory rules; if so, it triggers the constraint correction submodule. The rule engine's inference latency is less than 10 milliseconds, including approximately 5 milliseconds for rule matching and approximately 5 milliseconds for action correction.

[0037] To automatically find alternatives when actions violate constraints, the neural symbolic decision module also includes a constraint correction submodule. This submodule searches for the nearest available action when an RL action violates a constraint. The distance metric d(a,a') in the action space is defined as the Euclidean distance between the action parameters. The constraint correction algorithm searches the action space for an action a that satisfies all the mandatory rules, minimizing d(a_RL,a). The search employs a greedy strategy, first attempting to reduce the action magnitude (e.g., reducing Reduce load from 20% to 10%). If this still violates the constraint, it then tries adjacent actions (e.g., changing Reroute to Monitor). To address the extreme case where all candidate actions violate constraints, the neural symbolic decision-making module also includes a degradation mode trigger. This trigger switches to a conservative strategy when constraints cannot be met. Degradation modes include minimal intervention (only executing the Monitor action and awaiting manual decision-making), a safety mode (performing an Emergency shed to reduce load and ensure system safety), and an offline mode (disconnecting AI control and switching to a traditional protection and control system). The selection of the degradation mode is based on the severity of the fault and the system state, implemented through a pre-defined decision tree, providing a final safety net for the system.

[0038] In one embodiment, the fault recovery execution module is specifically used to execute the following phased process: During the fault detection phase, faults are identified and located based on abrupt changes in topology metrics and / or anomalies in consistency scores. During the fault isolation phase, the isolation priority of the faulty line is determined based on topology metrics and consistency scores, and the circuit breakers are coordinated to perform isolation actions. During the topology reconfiguration phase, optimization algorithms are used to search for reconfiguration paths and optimize power allocation on the remaining healthy lines. During the load recovery phase, the disconnected load is gradually restored, and a safety check is performed during the recovery process.

[0039] Specifically, the fault recovery execution module is used to coordinate circuit breaker operation, topology reconfiguration, and load balancing to achieve phased fault isolation and power restoration. The fault recovery execution module includes four phases: fault detection, fault isolation, topology reconfiguration, and load restoration.

[0040] (1) Fault detection stage: The fault detection phase achieves rapid fault identification through topology indicator mutation detection and data consistency anomaly detection. The system maintains a sliding window history of topology indicators (window length 10 periods, approximately 10 to 50 seconds), calculates the mean μ and standard deviation σ of the indicators, and triggers mutation detection when the current indicator value exceeds μ ± 3σ. Typical mutation patterns include a sudden drop in algebraic connectivity λ_2 (decrease greater than 30%), a sudden increase in the vulnerability index F (increase greater than 50%), and a sudden drop in consistency score (multiple buses simultaneously falling below 0.5). The fault detection latency is less than 2 seconds, including approximately 50 milliseconds for indicator calculation, approximately 100 milliseconds for mutation detection, and approximately 1 second for alarm triggering (including human-computer interaction confirmation).

[0041] (2) Fault isolation phase: During the fault isolation phase, the faulty line is rapidly isolated through coordinated operation of circuit breakers. Based on topology analysis and consistency checks, the system identifies a set of suspected faulty lines. For each suspected line Calculate isolation priority ,in Edge criticality (value range: 0 to 1). This is the average consistency score of the busbars at both ends of the line (range: -1 to 1). Higher priority indicates a more likely faulty line and a smaller impact from isolation. The system sends circuit breaker trip commands sequentially in descending priority order. After each isolation, the topology indicators are recalculated; isolation stops if the indicators return to normal. The circuit breaker operating time is 100 to 200 milliseconds, the coordination delay is approximately 2 to 5 seconds, and the total fault isolation time is less than 10 seconds.

[0042] (3) Topology Reconstruction Stage: The topology reconstruction phase achieves healthy topology reconstruction through ACO (Ant Colony Optimization) path discovery and PSO (Particle Swarm Optimization) load balancing. The ACO algorithm searches for low-congestion paths on the remaining healthy paths. Virtual ants select their next hop based on pheromone concentration and a heuristic function (1 / congestion level), depositing pheromones on low-congestion paths. After 50 to 100 iterations, it converges to the optimal path set. The PSO algorithm optimizes power allocation on feasible paths. Particles explore the load allocation space and converge to the optimal equilibrium point through velocity and position updates, satisfying generator output limits, ramp rate, and power balance constraints. The ACO+PSO hybrid optimization takes approximately 60 seconds, including ACO path discovery (approximately 30 seconds), PSO load balancing (approximately 20 seconds), and DC power flow verification (approximately 10 seconds).

[0043] (4) Load recovery phase: During the load restoration phase, the disconnected loads are restored gradually, prioritizing critical loads (hospitals, communications, transportation) and large user loads. Based on the topology reconfiguration results, the system calculates the available capacity margin for each bus and generates a restoration sequence according to load priority and capacity constraints. After each load group is restored, a DC power flow check is performed to verify whether voltage, frequency, and line utilization meet safety constraints. If the check fails, the system reverts to the previous state, adjusting the restoration sequence or reducing the restoration power. The load restoration employs a gradual strategy, restoring 10% to 20% of the disconnected load each time, observing the system response at 5 to 10-second intervals, with a total restoration time of approximately 50 seconds.

[0044] To ensure the system truly enters steady-state operation after recovery, in a preferred embodiment of the invention, the fault recovery execution module further includes a system stabilization phase. During this phase, voltage, frequency, and power oscillations are monitored to ensure the system enters steady-state operation. The system uses a sliding window monitoring method (window length 30 seconds) to calculate the standard deviation of voltage and frequency. System stability is determined when the standard deviation is less than a threshold (voltage 0.01 pu, frequency 0.05 Hz) for three consecutive windows. If continuous oscillation is detected (standard deviation greater than the threshold for more than 2 minutes), damping control is triggered or the topology is readjusted. The system stabilization time is approximately 50 seconds, including monitoring delay (approximately 30 seconds) and damping control (approximately 20 seconds).

[0045] like Figure 4 As shown, based on the same inventive concept, this embodiment of the invention also provides a method for rapid power grid fault recovery based on graph data detection, including: S1. Obtain the topology data of the power grid, construct a weighted graph representing the topology of the power grid, and calculate the spectral features of the weighted graph to generate a topology metric for quantifying the resilience of the power grid. S2. Acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and calculate the geometric consistency between each local measurement data and the global state data of the corresponding area to generate a consistency score to indicate data quality. S3. Based on the pre-trained reinforcement learning model, topological metrics and consistency scores are fused to generate high-level control intentions, and a symbol rule base with physical constraints is used to verify and correct the high-level control intentions to generate control actions. S4. Based on the control actions, generate and issue specific equipment control commands to coordinate circuit breaker actions, topology reconfiguration, and load balancing in the power grid, thereby achieving phased fault isolation and power restoration.

[0046] To verify the effectiveness of the power grid fault rapid recovery system and method based on graph data detection of the present invention, the following is an example verification on the IEEE 118-bus standard test system.

[0047] 1. Example 1: Verification of self-healing of N-1 faults This embodiment uses the IEEE 118-node system to verify the system's self-healing capability under an N-1 fault scenario. The IEEE 118-node system comprises 118 bus nodes, 186 transmission lines, 54 generators, and 91 load nodes, and is an internationally recognized standard test system for power grid simulation. The experimental configuration is that the system initially operates under normal conditions, with a total load of 4242 MW, total power generation of 4519 MW (including 277 MW of spinning reserve), and a renewable energy penetration rate of 25% (wind power accounting for 15% and photovoltaic power accounting for 10%). At t=10 seconds, an N-1 fault is manually triggered on line 30-38 (the transmission line connecting bus 30 and bus 38). This line is one of the critical lines of the system, and its fault will lead to a decrease in local topology connectivity and power flow redistribution.

[0048] (1) Fault detection phase (t=10 to 12 seconds): The topology analysis module completed the topology indicator calculation within 50 milliseconds after the fault occurred. The algebraic connectivity λ_2 suddenly dropped from the normal value of 0.18 to 0.11 (a decrease of 39%), and the vulnerability index F suddenly increased from the normal value of 28 to 47 (an increase of 68%), triggering a mutation detection alarm. The data consistency detection module detected that the consistency score of bus 30 and bus 38 suddenly dropped from the normal value of 0.85 to 0.32 (below the threshold of 0.5), marking them as suspected fault areas. The total fault detection delay was 1.8 seconds, including 50 milliseconds for indicator calculation, 100 milliseconds for mutation detection, and 1.65 seconds for human-computer interaction confirmation.

[0049] (2) Fault isolation phase (t=12 to 20 seconds): The system calculates the isolation priority of lines 30-38 as p_e = 0.78 × (1 - 0.32) = 0.53 (edge ​​criticality c_e = 0.78, consistency score consistency_e = 0.32), ranking it first in priority. The system sends a circuit breaker trip command, with a circuit breaker action time of 150 milliseconds. After isolation, the topology indicators are recalculated, λ_2 recovers to 0.15, F drops to 35, and the indicators return to normal. The total fault isolation time is 8 seconds, including 2 seconds for priority calculation, 5 seconds for circuit breaker coordination, and 1 second for indicator verification.

[0050] (3) Topology reconstruction phase (t=20 to 85 seconds): The ACO algorithm searches for low-congestion paths among the remaining 185 healthy lines, converging to the optimal path set after 80 iterations. It discovers that a detour via line 30-17-113-32-113-38 bypasses the faulty line. The PSO algorithm optimizes power allocation on feasible paths, converging to the optimal equilibrium point after 45 iterations, satisfying all generator output limits (minimum output 10 MW to maximum output 400 MW) and ramp rate limits (3% rated power / minute). DC power flow verification shows that all line utilization is less than 0.85, and the voltage range is 0.96 to 1.04 pu, meeting safety constraints. The total topology reconfiguration time is 65 seconds, including 35 seconds for ACO path discovery, 20 seconds for PSO load balancing, and 10 seconds for DC power flow verification.

[0051] (4) Load recovery phase (t=85 to 135 seconds): The system gradually restores the disconnected loads according to priority (120 MW of load near bus 30 was disconnected during fault isolation). First, 40 MW of primary load (hospital and telecommunications) is restored, and the system response is observed after 10 seconds; voltage and frequency are stable. Then, 50 MW of secondary load (large users) is restored, and the line utilization rate is verified after 10 seconds; it is not exceeded. Finally, 30 MW of tertiary load (residential) is restored, with a total restoration time of 50 seconds.

[0052] (5) System stability phase (t=135 to 185 seconds): The system monitors the sliding window standard deviations of voltage and frequency. The voltage standard deviation gradually decreases from an initial value of 0.025 pu to 0.008 pu (less than the threshold of 0.01 pu), and the frequency standard deviation decreases from 0.12 Hz to 0.03 Hz (less than the threshold of 0.05 Hz). The system is considered stable if the stability condition is met for three consecutive windows (30 seconds each). The total stabilization time is 50 seconds.

[0053] Experimental results show that, under the N-1 fault scenario, the system achieves fault detection in 1.8 seconds, fault isolation in 8 seconds, topology reconstruction in 65 seconds, load recovery in 50 seconds, and system stabilization in 50 seconds, with a total recovery time of approximately 175 seconds (about 2.9 minutes), meeting the design target of 2 to 3 minutes in AI self-healing mode. Compared to traditional protection and control systems (which require 12 to 15 minutes for manual decision-making), the recovery time is reduced by approximately 80%.

[0054] 2. Example 2: Verification of N-2 Fault Cascade Failure Defense This embodiment verifies the system's ability to prevent cascading failures under an N-2 fault scenario. An N-2 fault refers to the simultaneous or successive failure of two lines, posing a serious challenge to the power grid and potentially causing cascading failures and widespread power outages.

[0055] Experimental configuration: The system initially operates under high load conditions, with a total load of 4800 MW (load factor 95%) and a renewable energy penetration rate of 35% (wind power 20%, photovoltaic 15%). Faults on lines 30-38 are triggered at t=10 seconds, and faults on lines 23-24 are triggered at t=25 seconds (15 seconds after the first fault). Both faulted lines are critical to the system, and their simultaneous failure would lead to localized islanding risk.

[0056] (1) Fault detection stage: After the first fault (lines 30-38) was triggered, the algebraic connectivity λ_2 decreased from 0.16 to 0.09 (a 44% decrease), and the vulnerability index F increased from 32 to 58 (an 81% increase), triggering an alarm. After the second fault (lines 23-24) was triggered, λ_2 further decreased to 0.03 (approaching the islanding threshold of 0.05), and F increased to 95 (exceeding the vulnerability threshold of 50), leading the system to determine a high-risk state. The data consistency detection module detected that the consistency scores of four buses (buses 23, 24, 30, and 38) were below 0.5, marking them as suspected fault areas. The total fault detection delay was 2.2 seconds (including the cumulative detection time of the two faults).

[0057] (2) Fault isolation phase: The system calculates the isolation priority of the two faulty lines: lines 30-38 have a priority of 0.53, and lines 23-24 have a priority of 0.61 (the edges are more critical). Lines are isolated in descending order of priority. After isolating lines 23-24, λ_2 recovers to 0.08, and F drops to 62, but is still above normal levels. After isolating lines 30-38, λ_2 recovers to 0.12, and F drops to 42, with the indicators basically recovered. The total fault isolation time is 12 seconds (6 seconds for each of the two isolations).

[0058] (3) Topology Reconstruction Stage: The ACO algorithm searches for low-congestion paths bypassing the two faulty lines. Due to the reduced number of feasible paths, the search space increases, leading to 120 iterations and a convergence time of 40 seconds. The PSO algorithm optimizes power allocation. Due to tighter constraints (two line faults reducing available capacity), the number of iterations increases to 60, with a convergence time of 25 seconds. DC power flow verification shows that the utilization rate of some lines reaches 0.88 (close to the upper limit of 0.9), requiring the shedding of some low-priority loads to meet safety constraints. The total topology reconfiguration time is 75 seconds.

[0059] (4) Load recovery phase: The system disconnected 180 MW of tertiary loads to meet safety constraints, and then restored tertiary and secondary loads according to priority. Due to limited available capacity, only 60 MW of tertiary loads and part of the 40 MW of secondary loads were restored, while 80 MW of tertiary loads remained disconnected, awaiting restoration after the faulty line was repaired. The total load restoration time was 60 seconds.

[0060] (5) System stabilization phase: Due to the significant load shedding, the system stabilization time was reduced to 40 seconds. Ultimately, the system achieved a total recovery time of approximately 190 seconds (about 3.2 minutes) under the N-2 fault scenario, meeting the design target of 3 to 5 minutes in the AI ​​self-healing mode.

[0061] Experimental results show that the system can effectively defend against cascading failures in the N-2 fault scenario. By quickly isolating faulty lines, reconstructing a healthy topology, and disconnecting low-priority loads, it ensures the safe and stable operation of the system. Compared with traditional methods (N-2 faults typically lead to large-scale power outages with recovery times ranging from hours to days), this invention significantly improves grid resilience.

[0062] 3. Example 3: Performance Comparison and Verification This embodiment verifies the effectiveness of the system through comparative experiments. The baseline comparison includes a traditional protection and control system (manual decision-making), a rule-based expert system, and the complete system of this invention. The experiment is configured to perform 100 random fault simulations on an IEEE 118-bus system, including 70 N-1 faults, 20 N-2 faults, and 10 load surges. Evaluation metrics include fault detection delay, fault isolation time, topology reconfiguration time, total recovery time, constraint violation rate, load loss, and system stability.

[0063] (1) Performance comparison results: Performance comparison results show that the traditional protection and control system has an average recovery time of 720 seconds (12 minutes), a constraint violation rate of 5% (mainly due to human decision-making errors), and an average load loss of 180 MW. The rule-based expert system has an average recovery time of 240 seconds (4 minutes), a constraint violation rate of 8% (due to incomplete rule base coverage), and an average load loss of 150 MW. The complete system of this invention has an average recovery time of 165 seconds (2.75 minutes), a constraint violation rate of only 2.5% (due to effective MPC constraint optimization), and an average load loss of 100 MW.

[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A rapid power grid fault recovery system based on graph data detection, characterized in that, include: The topology analysis module is used to acquire the topology data of the power grid, construct a weighted graph characterizing the topology of the power grid, and calculate the spectral features of the weighted graph to generate a topology metric for quantifying the resilience of the power grid. The data consistency detection module is used to acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and to calculate the geometric consistency between each local measurement data and the global state data of the corresponding area to generate a consistency score to indicate data quality. The neural symbolic decision module is used to receive the topology metric and the consistency score, fuse them based on a pre-trained reinforcement learning model to generate a high-level control intention, and use a symbolic rule library encoded with physical constraints to verify and correct the high-level control intention to generate control actions. The fault recovery execution module is used to generate and issue specific equipment control commands based on the control actions, so as to coordinate the operation of circuit breakers, topology reconfiguration and load balancing in the power grid, and realize phased fault isolation and power supply restoration.

2. The system according to claim 1, characterized in that, The topology analysis module is further used to calculate the eigenvalues ​​of the Laplacian matrix of the weighted graph, and the second smallest eigenvalue is used as an algebraic connectivity index to quantify the risk of grid splitting.

3. The system according to claim 1, characterized in that, The data consistency detection module is further used to map the local measurement data and the global state data into vectors, and obtain the consistency score by calculating the normalized inner product between the vectors.

4. The system according to claim 1, characterized in that, The neural symbol decision-making module further includes: The RL agent submodule is used to output the higher-level control intent based on the current system state; A symbol rule library is used to store if-then rules that encode at least one physical constraint among generator ramp rate, circuit breaker operating time, and transformer tap change interval. The rules engine is used to trigger the correction of the higher-level control intent or trigger the system to enter a degraded operation mode when the higher-level control intent violates the rules in the symbol rule base.

5. The system according to claim 1, characterized in that, The fault recovery execution module is specifically used to execute the following phased process: During the fault detection phase, faults are identified and located based on abrupt changes in the topology metric and / or anomalies in the consistency score. During the fault isolation phase, based on the topology metric and the consistency score, the isolation priority of the faulty line is determined, and the circuit breaker is coordinated to perform isolation actions. During the topology reconfiguration phase, optimization algorithms are used to search for reconfiguration paths and optimize power allocation on the remaining healthy lines. During the load recovery phase, the disconnected load is gradually restored, and a safety check is performed during the recovery process.

6. A method for rapid recovery of power grid faults based on graph data detection, applied to the system as described in any one of claims 1 to 5, characterized in that, Includes the following steps: S1. Obtain the topology data of the power grid, construct a weighted graph representing the topology of the power grid, and calculate the spectral features of the weighted graph to generate a topology metric for quantifying the resilience of the power grid. S2. Acquire local measurement data from multiple measuring points in the power grid, as well as global state data obtained based on state estimation, and calculate the geometric consistency between each of the local measurement data and the global state data of the corresponding region to generate a consistency score for indicating data quality. S3. Based on the pre-trained reinforcement learning model, the topology metric and the consistency score are fused to generate a high-level control intention, and the high-level control intention is verified and corrected using a symbol rule base encoded with physical constraints to generate control actions. S4. Based on the control actions, generate and issue specific equipment control commands to coordinate circuit breaker actions, topology reconfiguration, and load balancing in the power grid, thereby achieving phased fault isolation and power supply restoration.

Images