Hydropower station direct current system fault detection and elimination method and system based on heterogeneous graph and multi-agent

Abstract

The invention discloses a hydropower station direct current system fault detection and elimination method and system based on a heterogeneous graph and multiple agents, and belongs to the field of power system fault diagnosis. In order to solve the problems of incomplete fault perception, low positioning precision, insufficient collaborative decision and poor execution security of an existing method, multi-source heterogeneous data is collected and subjected to edge side fusion processing, and a heterogeneous graph digital twinborn body containing node types, edge relations and meta-paths is constructed; a heterogeneous graph attention network model is adopted, and accurate fault positioning is achieved through node-level and semantic-level attention learning; generating a collaborative self-healing strategy based on a multi-agent depth deterministic strategy gradient algorithm, and issuing and executing the collaborative self-healing strategy through a GOOSE protocol after a digital twin experience certificate meets security constraints; and a closed-loop optimization mechanism is constructed in combination with the block chain. According to the method, accurate coupling fault identification, multi-partition cooperative self-healing and safe and rapid execution are realized, the fault processing time is shortened from an hour level to a minute level, and the operation reliability and the intelligent level of a direct current system of a hydropower station are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system fault diagnosis technology, and in particular to a method and system for fault detection and elimination in hydropower station DC systems based on heterogeneous graphs and multi-agent systems. Background Technology

[0002] Hydropower stations, as the core pillar of the clean energy supply system, bear the critical missions of power generation, grid frequency regulation and peak shaving, and emergency power supply. The DC system, as the "lifeline" power source for the hydropower station's control circuits, relay protection devices, communication equipment, and emergency lighting, directly determines the core capabilities of unit start-up and shutdown, fault isolation, and grid safety and stability. With the expansion of new energy grid integration and the increasing complexity of grid operation, hydropower stations are placing more stringent demands on the continuous power supply, fault response speed, and self-healing capabilities of their DC systems. A DC system failure could lead to unplanned unit shutdowns, failure or malfunction of critical protection systems, and consequently trigger cascading grid failures, causing significant economic losses and social impact.

[0003] From the perspective of the system's inherent characteristics, the DC system of a hydropower station exhibits significant features of "heterogeneity, complexity, and strong coupling": it includes a variety of heterogeneous devices with significantly different functions, such as battery banks (up to hundreds of individual units), high-frequency charging devices, multi-segment DC buses, dozens to hundreds of DC feeders, insulation monitoring devices, and contactors. The types of nodes are numerous and the connection relationships are complex (covering multiple relationships such as electrical connections, power supply dependence, control command transmission, and physical proximity). At the same time, the operating environment of hydropower stations is harsh. High humidity can easily lead to insulation aging of equipment, strong electromagnetic interference can affect the accuracy of monitoring signals, and mechanical vibration can accelerate the wear of contactor contacts and loosening of battery racks. These factors together lead to the DC system's fault modes exhibiting characteristics such as concealment (e.g., slow increase in battery internal resistance, gradual increase in charging module ripple coefficient exceeding the standard), coupling (single equipment failure triggers multiple feeder power supply anomalies), and propagation (faults spread from the power supply end to the control end). Traditional detection methods are difficult to fully cover these characteristics.

[0004] Currently, the industry still relies on traditional technical approaches for fault detection and troubleshooting of DC systems in hydropower stations, which have many limitations that are difficult to overcome: First, the detection dimensions are singular and one-sided. Existing solutions focus on insulation monitoring, mainly using methods such as bridge method and leakage current monitoring to focus on the problem of insulation resistance decline. However, they lack effective detection capabilities for high-frequency faults in the DC system of hydropower stations (such as battery cell capacity decay, excessive internal resistance, abnormal output ripple of charging device, mechanical jamming of contactor, and overheating of feeder connectors). This leads to the long-term accumulation of a large number of "hidden faults" and eventually causes sudden system failure. Secondly, the fault location accuracy is low. In complex wiring scenarios with multiple feeders and branches, traditional insulation monitoring devices can only locate the general area where the fault is located (such as a section of busbar or a feeder cabinet), and cannot accurately pinpoint the faulty equipment or specific node. It is necessary to rely on maintenance personnel to disconnect the power line by line to check, which is not only time-consuming and labor-intensive (a single check usually takes 2-3 hours), but may also expand the scope of the fault due to misoperation. Third, the lack of an intelligent collaborative decision-making mechanism means that fault handling relies heavily on the experience and judgment of maintenance personnel, resulting in slow response times (often exceeding hours from the occurrence of a fault to its completion). Furthermore, it cannot automatically generate a globally optimal system reconfiguration strategy based on the real-time status of the system (such as the load distribution of each zone, the capacity of the backup power supply, and the health status of the equipment), which can easily lead to the problem of "local processing taking precedence and global power supply imbalance." For example, after a single zone fault is isolated, the backup power supply is not activated in time, causing the associated zone protection device to lose power. Fourth, existing intelligent solutions have technical shortcomings. Some related research, such as the intelligent fault location system and method based on DC systems disclosed in CN120652198A, although it introduces digital twin and graph neural network technologies, treats heterogeneous devices with different functions and characteristics (such as batteries and chargers, feeders and protection devices) as nodes of the same type, ignoring the semantic specificity of device functional differences and connection relationships, resulting in inaccurate feature extraction and one-sided fault correlation analysis. At the same time, the single-agent reinforcement learning strategy it adopts is difficult to adapt to the complex scenario of multiple power supply zones and multiple feeder coordination in large hydropower stations, and cannot realize the linkage decision-making of devices in each zone. The strategy optimization is limited to the local and it is difficult to achieve the goal of maximizing global power supply reliability.

[0005] From an industry development perspective, hydropower stations are rapidly evolving towards larger scale (installed capacity exceeding 1000MW) and greater intelligence (unmanned or minimally staffed operation). The number of feeders, equipment size, and operational complexity of DC systems are continuously increasing, and traditional technologies are no longer sufficient to meet the intelligent operation and maintenance requirements of "early fault warning, precise location, second-level self-healing, and safe execution." Furthermore, as the power grid's assessment of unplanned outages at hydropower stations becomes increasingly stringent, each unit outage caused by the DC system results in direct economic losses of millions of yuan, while also affecting the stability of regional power supply. Therefore, breaking through existing technological bottlenecks and constructing an intelligent solution that adapts to the heterogeneous characteristics of hydropower station DC systems, supports multi-regional collaboration, and possesses both precise detection and safe self-healing capabilities has become a critical technical issue that the industry urgently needs to address. This not only has significant technological innovation value but also substantial economic and social significance.

[0006] Based on this, the present invention proposes an integrated solution that combines heterogeneous graph neural networks, multi-agent reinforcement learning and digital twin technology. It specifically addresses the core pain points in existing technologies, such as inaccurate modeling of heterogeneous equipment, narrow fault detection dimensions, weak collaborative decision-making capabilities, and unsafe strategy execution. This solution enables early warning, accurate location, collaborative self-healing and safe execution of faults in the DC system of hydropower stations, providing core technical support for the safe and stable operation of hydropower stations and power grids. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method and system for fault detection and elimination of DC systems in hydropower stations based on heterogeneous graphs and multi-agents, which solves the three core problems in the field of fault detection of DC systems in hydropower stations: incomplete fault feature perception, weak ability to model heterogeneous equipment relationships, and low efficiency of multi-fault collaborative decision-making. At the same time, it overcomes the limitations of existing technologies such as single detection dimension, low positioning accuracy, and lack of intelligent decision-making.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method and system for fault detection and troubleshooting in a hydropower station DC system based on heterogeneous graphs and multi-agent architecture are provided, specifically including: 1. Fault Detection and Troubleshooting Method for DC Systems in Hydropower Stations Based on Heterogeneous Graphs and Multi-Agent Systems The method includes the following steps: Step 1: Collect multi-source heterogeneous data and perform edge-side fusion processing Collect multi-source heterogeneous data of the DC system of the hydropower station. The data includes the voltage and internal resistance of individual battery cells, the output ripple of the charging module, the leakage current of each DC feeder branch, the temperature of key nodes, the vibration signal of the switchgear, and the raw data of the insulation monitoring device.

[0009] Edge-side fusion processing is divided into data preprocessing and feature fusion: Data preprocessing: The ensemble empirical mode decomposition method is used to process the switchgear vibration signal and extract the intrinsic mode functions related to mechanical loosening; adaptive Kalman filtering is used to smooth the battery internal resistance measurement value to suppress the interference caused by periodic float charging / equalizing charging switching.

[0010] Feature fusion: Based on the knowledge graph, the preprocessed data is associated with the equipment ledger, topological relationships, and historical maintenance records to construct a triple of node-relationship-attribute, forming a unified feature representation.

[0011] Step 2: Construct a heterogeneous graph digital twin The construction process of a heterogeneous graph digital twin is as follows: Step 2.1: Define the node types of the heterogeneous graph, including batteries, chargers, AC power supplies, DC feeders, and loads (such as protection devices and controllers).

[0012] Step 2.2: Define the edge relationship types of the heterogeneous graph, including electrical connection, control relationship, physical proximity, and logical dependency.

[0013] Step 2.3: Based on the semantic enrichment of node features of the metapath, the metapath includes "battery - (power supply) -> DC feeder - (power supply) -> protection device". Information is aggregated through this metapath to enhance the perception of the risk of power failure of the protection device.

[0014] Step 3: Fault detection and localization based on heterogeneous graph neural networks The Heterogeneous Graph Attention Network (HAN) model is used to perform fault detection and localization. The specific steps are as follows: Step 3.1: Decompose the heterogeneous graph into multiple isomorphic subgraphs according to the preset meta-path (such as "device-connection-device" or "device-type-device").

[0015] Step 3.2: Perform node-level attention learning on each isomorphic subgraph, calculating the attention distribution of a node among its neighbors. For metapaths... ,node and nodes Attention coefficient The calculation is as follows: (1); In the formula, , , These are nodes , , initial characteristics, Metapath A specific weight matrix, It is an attention vector. It is the LeakyReLU activation function. It is a node In metapath The set of neighbors below.

[0016] Step 3.3: Perform semantic-level attention learning, fuse node embeddings from different meta-paths, and obtain the final node representation. The calculation formula is: (2); In the formula, Metapath The weight represents the importance of the metapath to the fault detection task. Metapath At the node The representation of; It is a set of metapaths.

[0017] Step 3.4: Input the final node representation into the classifier to predict the fault type (such as "battery failure", "insulation abnormality", "contactor jamming") and fault probability of each node.

[0018] Step 4: Generate a collaborative self-healing strategy based on multi-agent reinforcement learning Step 4.1: Divide the DC system into multiple intelligent agents according to the power supply zone, including the generator DC section Agent, the common section Agent, and the backup battery group Agent.

[0019] Step 4.2: Define the Markov game process for multi-agent cooperation: State space S: The joint local observations perceived by all agents, including the voltage of this zone, insulation state, load importance level, and state of adjacent zones.

[0020] Action Space A: The actions of each agent include load shedding, requesting support, isolating faults, and switching on backup power.

[0021] Reward function: includes global reward, local reward, and penalty term, where, Global Rewards If the entire DC system remains stable, a positive reward is given. Conversely, if a voltage loss occurs, a negative reward is given. The purpose of this reward is to ensure the stability of the entire system. If the entire DC system remains stable, a global positive reward is given. If pressure is lost, a negative reward will be given. ; Local rewards A local positive reward is given when the core protection device of this zone is continuously powered; a negative reward is given when the protection device is de-powered. If the core protection device of this zone is continuously powered, a local reward is given. ;otherwise ; Penalty items Penalize the mistaken shelving of non-critical loads, as this can lead to unnecessary system failures. The penalty can be quantified based on the severity of the mistaken shelving. If a non-critical load is mistakenly shelved, a penalty should be imposed. ; Combinatorial reward function : The state of the system at any given time At that time, perform the action The resulting comprehensive reward and punishment outcome; the final reward function can be expressed as: (3); In the formula, express The current state of the system (such as the current operating status of the DC system, load distribution, etc.); express Actions performed by the system at all times (such as load shedding, power supply adjustment, etc.); This indicates a global reward, which is given or penalized based on the stability of the entire station's DC system. This indicates a localized reward, which is given based on the operational status of the core protection devices in this zone. These are penalties, primarily targeting negative behaviors such as erroneous load shedding; Global rewards The weighting coefficients, For local rewards The weighting coefficients, For penalty items The weighting coefficients can be adjusted according to actual needs.

[0022] Step 4.3: Training is performed using the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. Each agent has an Actor network and a Critic network: the Actor network selects actions based on its own observations; the Critic network evaluates the value of the joint observations and joint actions of all agents and updates it by minimizing the temporal difference error. The loss function is: (4); in, (5); In the formula, Indicates the first The loss function of an agent-based Critic network; For the first Trainable parameters (weights, biases, etc.) of an agent's Critic network. For the mathematical expectation operator, the average is calculated based on samples in the empirical replay buffer D; For the first The current evaluation value (state-action value function) of the Critic network for each agent; For the joint observation of all agents (an instance of state space S); Indicates all The joint action of several agents (an instance of action space A); For timely rewards after the joint operation is carried out (rewards) (output) The next joint observation (next state) after the joint action is executed; The time-series difference target value (an unbiased estimate of the true value); For the first Immediate rewards for actions performed jointly by agents in a Critic network; Discount factor ( ); Indicates the first The target network evaluation value of the agent Critic network (evaluating the future value of "next joint observation + next joint action"); This represents the joint action of all agents in the next state (generated by the Actor network). Indicates the next state The joint action of several intelligent agents; Indicates the first The Actor target network policy function for each agent; For all Local observation of an agent.

[0023] Step 4.4: After training is completed, each agent shares intent through a communication mechanism based on real-time observation of the distributed execution strategy, thereby achieving collaborative fault isolation and power restoration.

[0024] Step 5: Verify and securely execute the collaborative self-healing strategy based on a digital twin. Step 5.1: Simulate the execution of the collaborative self-healing strategy in the digital twin to predict the dynamic response of the system.

[0025] Step 5.2: Perform a security check to verify whether the policy meets the hard constraints such as "N-1 power supply safety principle" and "uninterrupted core protection power supply". If the check fails, the policy rollback is triggered and a conservative predefined policy is adopted.

[0026] Step 5.3: After successful verification, control commands are sent to devices such as smart circuit breakers and contactors via the GOOSE (General Object-Oriented Substation Events) protocol to complete fault isolation and network reconfiguration.

[0027] Step 6: Construct a closed-loop optimization mechanism Step 6.1: Record all data from the entire fault handling process (including detection data, decision data, and execution result data) to the blockchain to ensure that the data is tamper-proof and to be used for post-event analysis and accountability.

[0028] Step 6.2: Utilize the accumulated field data to periodically fine-tune the heterogeneous graph neural network model and multi-agent reinforcement learning strategy offline, thereby achieving continuous evolution of the system.

[0029] 2. Fault Detection and Troubleshooting System for Hydropower Station DC Systems Based on Heterogeneous Graphs and Multi-Agents This system is used to implement the above methods and includes a multi-source heterogeneous graph construction module, a digital twin construction module, a heterogeneous neural network module, a strategy module, and a blockchain module. These modules are interconnected and can transmit and share data. Multi-source heterogeneous graph construction module: This module includes a sensor network module and a heterogeneous graph construction module. The sensor network module is used to collect multi-source operating data of the hydropower station's DC system and transmit the data to the heterogeneous graph construction module. The heterogeneous graph construction module is used to perform structured processing on the multi-source operating data to form a multi-source heterogeneous data structure, and to construct a heterogeneous graph based on node type, edge relationship type, and meta-path.

[0030] Digital Twin Construction Module: This module includes a digital twin module and a bidirectional real-time synchronization interface. The bidirectional real-time synchronization interface enables bidirectional data synchronization of the operational status between the constructed digital twin and the physical entity of the hydropower station's DC system, providing a virtual environment for strategy simulation and verification.

[0031] The heterogeneous neural network module includes a data input module and a fault detection and localization module. The data input module is used to input the structured multi-source heterogeneous data output by the multi-source heterogeneous graph construction module into the fault detection and localization module; the fault detection and localization module adopts the heterogeneous graph attention network (HAN) model to perform fault detection and localization operations and output the fault type and fault probability.

[0032] The strategy module includes a multi-agent reinforcement learning model and a digital twin strategy verification module. The multi-agent reinforcement learning model communicates with the heterogeneous neural network module and generates a collaborative self-healing strategy based on fault detection and localization results. The digital twin strategy verification module is connected to both the multi-agent reinforcement learning model and the digital twin construction module to simulate the execution of the collaborative self-healing strategy in the digital twin and verify the feasibility of the strategy.

[0033] Blockchain module: Used to record and store data throughout the entire fault handling process, ensuring that the data is tamper-proof and providing data support for closed-loop system optimization.

[0034] The fault detection and troubleshooting method and system for hydropower station DC systems based on heterogeneous graphs and multi-agents provided by this invention have the following beneficial effects: 1. This invention successfully overcomes the shortcomings of existing technologies, such as incomplete fault feature perception, weak heterogeneous equipment relationship modeling capability, low efficiency of multi-fault collaborative decision-making, and insufficient security of strategy execution. It provides a method and system for fault detection and elimination of hydropower station DC systems that can accurately identify multiple types of coupled faults, quickly generate collaborative self-healing strategies, safely execute fault handling operations, and be continuously optimized.

[0035] 2. Improved Fault Detection Precision: This invention uses a heterogeneous graph neural network to finely model the differences and complex relationships between heterogeneous devices. Combined with the fusion of multi-source heterogeneous data, it achieves high-precision location of coupled faults and latent faults. This effectively solves the problems of single detection dimensions and ambiguous location in traditional detection methods. It can comprehensively detect multiple types of faults such as increased battery internal resistance, abnormal charging modules, and aging contactor contacts.

[0036] 3. Decision-making collaboration optimization: This invention adopts a multi-agent reinforcement learning algorithm, which divides agents according to power supply zones and makes collaborative decisions. It overcomes the limitations of single-agent decision-making perspective, generates a globally optimal self-healing strategy, improves the processing efficiency in multi-fault scenarios, and realizes collaborative decision-making for multi-zone power supply.

[0037] 4. Enhanced safety and reliability: This invention uses a digital twin for strategy simulation and safety verification, strictly adheres to the N-1 power supply safety principle and the constraint of uninterrupted core protection power supply, avoids the expansion of faults caused by misoperation, ensures the safe execution of control commands, and effectively prevents the expansion of faults due to misoperation.

[0038] 5. Rapid Fault Recovery: The fully automated and intelligent processing of this invention reduces fault handling time from "hours" to "minutes" compared to traditional manual methods, significantly improving the power supply reliability of the DC system in hydropower stations.

[0039] 6. Continuous System Evolution: This invention adopts a blockchain-based immutable data storage and offline fine-tuning mechanism to achieve continuous optimization of models and strategies, adapting to scenarios such as equipment aging and changes in operating conditions during the long-term operation of hydropower station DC systems.

[0040] 7. Enhanced Equipment Relationship Modeling Capability: This invention introduces a heterogeneous graph neural network model to replace the traditional homogeneous graph neural network, effectively distinguishing different types of equipment and their complex connection relationships in the DC system of a hydropower station, and improving the ability to perceive fault characteristics.

[0041] 8. Solving the problem of insulation monitoring and positioning: This invention successfully overcomes the shortcomings of traditional insulation monitoring devices, such as unclear positioning in complex branch situations, reliance on manual inspection of each branch, resulting in low positioning accuracy and low efficiency.

[0042] 9. Eliminating reliance on human experience: This invention successfully overcomes the limitations of fault handling relying on the experience of operation and maintenance personnel, slow response speed, and inability to automatically generate and execute optimal system reconfiguration strategies, thereby avoiding the expansion of power outage scope and ensuring safe shutdown of the unit.

[0043] 10. Overcoming the shortcomings of single agent: This invention successfully overcomes the problem that single agent reinforcement learning is insufficient in terms of policy coordination and optimality when dealing with complex DC networks with multiple zones and feeders in hydropower stations.

[0044] 11. Advantages of data preprocessing and fusion: This invention uses edge computing nodes for data preprocessing and feature fusion, which solves the problems of single detection dimension and low positioning accuracy of traditional methods, and improves the comprehensiveness and accuracy of fault detection.

[0045] 12. Precisely perceive equipment differences: This invention uses a heterogeneous graph neural network to finely model the differences and complex relationships among various equipment in the DC system of a hydropower station, enabling high-precision location of coupled faults and latent faults.

[0046] 13. Collaborative Decision-Making Perspective Optimization: This invention adopts multi-agent reinforcement learning to overcome the shortcomings of single agents having a single decision-making perspective in complex systems, realizes collaborative assistance among power supply zones, and generates a globally better self-healing strategy.

[0047] 14. Security Verification Ensures Safety: This invention uses a digital twin for strategy simulation and security verification, effectively preventing the escalation of faults caused by misoperation and ensuring the absolute safety of control commands.

[0048] 15. Intelligent Upgrade of the Entire Process: This invention automates and intelligently upgrades the entire process from fault detection and location to strategy generation and execution, reducing fault handling time from "hours" to "minutes" in traditional manual handling, and greatly improving power supply reliability.

[0049] 16. The beneficial effects of the present invention have been fully verified through theoretical derivation and practical application cases, demonstrating the significant advantages of the present invention in improving the intelligent level of fault handling and operational reliability of DC systems in hydropower stations. Attached Figure Description

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the technical process of the method of the present invention. Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0051] The technical solutions of the present invention will be further described below with reference to the embodiments and accompanying drawings: Example 1 This embodiment provides a method for fault detection and troubleshooting of DC systems in hydropower stations based on heterogeneous graphs and multi-agent systems, combining... Figure 1 The process shown is implemented as follows: 1. Application Scenarios This embodiment is applied to a large hydropower station with an installed capacity of 1200MW. Its DC system adopts a double busbar segmented connection to provide control, protection and emergency lighting power for four 300MW units. It includes four sets of batteries, eight charging devices and 56 DC feeder cabinets.

[0052] 2. System Configuration (1) Sensor Networks: Each battery cell is equipped with a voltage / internal resistance monitoring unit (sampling rate 1Hz, accuracy ±0.5%); the charging device is equipped with a ripple acquisition unit (threshold ≤3%); each DC feeder branch is deployed with a leakage current sensor with a resolution of 0.1mA; temperature sensors are installed in the battery room, charging device cabinet, etc.; and acceleration sensors are deployed on the DC power supply cabinet door and battery rack.

[0053] (2) Edge computing nodes: Eight Intel XeonD-2145NT industrial-grade edge servers (128GB RAM) were deployed, each responsible for data processing and model inference in one power supply zone.

[0054] Communication network: The station control layer adopts the MMS (Manufacturing Message Specification) protocol (transmission rate 100Mbps), and the process layer adopts the GOOSE (Generic Object Oriented Substation Events) protocol (transmission delay <4ms).

[0055] 3. Specific implementation process like Figure 1 As shown, the implementation process of this method is as follows: Step 1: Multi-source heterogeneous data acquisition and edge-side fusion processing Data were collected on individual battery cell voltage (normal range 2.23~2.28V), internal resistance (normal range 0.5~1.2mΩ), temperature (15~30℃), as well as charging device ripple, feeder leakage current, and equipment vibration. When the internal resistance of battery pack #3 increased from 0.8mΩ to 1.5mΩ, the edge nodes used adaptive Kalman filtering to smooth the data and confirm the authenticity of the abnormal internal resistance trend.

[0056] Step 2: Construct a heterogeneous graph digital twin Generate a heterogeneous graph containing 328 nodes and 756 edges: node types cover "battery, charger, DC power supply, feeder, protection device, control system"; edge relationships are defined as "electrical connection, power supply relationship, control relationship, physical location"; meta-paths are set as "battery-power supply -> protection device" and "charger-charging -> battery", and a digital twin is formed based on real-time data synchronization.

[0057] Step 3: Fault Detection in Heterogeneous Graph Neural Networks Processing heterogeneous graph data using a Heterogeneous Graph Attention Network (HAN): First, decompose the heterogeneous graph into homogeneous subgraphs according to the meta-paths "battery-power supply -> protection device" and "charger-charging -> battery". Node-level attention learning: Calculate the attention coefficient of node pairs using the following formula: (1); In the formula, Metapath Middle node and nodes Attention coefficient; , , These are nodes , , initial characteristics, Metapath A specific weight matrix, It is an attention vector. It is the Leaky ReLU activation function. It is a node In metapath The set of neighbors below.

[0058] Semantic attention learning: The final node representation is obtained by fusing features from multiple subgraphs, as shown in the formula: (2); In the formula, This represents the final node representation; Metapath The weight represents the importance of the metapath to the fault detection task. Metapath At the node The representation of; It is a set of metapaths.

[0059] Ultimately, the failure probability of the #3 battery pack was found to be 92% (abnormal internal resistance), and the power supply risk of the #1 unit protection system was also identified (probability 65%).

[0060] Step 4: Multi-agent reinforcement learning decision making The system is divided into four agent partitions, and policies are collaboratively generated using the MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm. The reward function in the algorithm includes a global reward, a local reward, and a penalty term. Global Rewards If the entire DC system remains stable, a positive reward is given. Conversely, if a voltage loss occurs, a negative reward is given. The purpose of this reward is to ensure the stability of the entire system. If the entire DC system remains stable, a global positive reward is given. If pressure is lost, a negative reward will be given. ; Local rewards A local positive reward is given when the core protection device of this zone is continuously powered; a negative reward is given when the protection device is de-powered. If the core protection device of this zone is continuously powered, a local reward is given. ;otherwise ; Penalty items Penalize the mistaken shelving of non-critical loads, as this can lead to unnecessary system failures. The penalty can be quantified based on the severity of the mistaken shelving. If a non-critical load is mistakenly shelved, a penalty should be imposed. ; Combinatorial reward function : The state of the system at any given time At that time, perform the action The resulting comprehensive reward and punishment outcome; the final reward function can be expressed as: (3); In the formula, express The current state of the system (such as the current operating status of the DC system, load distribution, etc.); express Actions performed by the system at all times (such as load shedding, power supply adjustment, etc.); This indicates a global reward, which is given or penalized based on the stability of the entire station's DC system. This indicates a localized reward, which is given based on the operational status of the core protection devices in this zone. These are penalties, primarily targeting negative behaviors such as erroneous load shedding; Global rewards The weighting coefficients, For local rewards The weighting coefficients, For penalty items The weighting coefficients can be adjusted according to actual needs.

[0061] The loss function of the Critic network in the algorithm is: (6); The target value for time-series difference is: (7); In the formula, Corresponding to 4 agent serial numbers; Indicates the first The loss function of an agent-based Critic network; For the first Trainable parameters of an agent Critic network; For the mathematical expectation operator, the average is calculated based on samples in the empirical replay buffer D; For the first The current evaluation value (state-action value function) of the Critic network for each agent; For the joint observation of "internal resistance of #3 battery pack + protection voltage of #1 unit"; This is indicated by combined actions such as "requesting support" and "switching to a backup battery." An immediate reward for "Core Protection Not Losing Power" (set to +80); The next joint observation (next state) after the joint action is executed; The time-series difference target value (an unbiased estimate of the true value); For the first Timely rewards for joint actions executed by individual agents in the Critic network; Set as a discount factor (0.9); Indicates the first The target network evaluation value of the agent Critic network (evaluating the future value of "next joint observation + next joint action"); This represents the joint action of all agents in the next state (generated by the Actor network). Indicates the next state The joint action of several intelligent agents; Indicates all A target network jointly observed by multiple intelligent agents; For all Joint observation by multiple agents.

[0062] The final strategy is to complete the isolation of battery pack #3 and the activation of the backup battery pack within 30 seconds.

[0063] Step 5: Digital Twin Policy Verification and Security Execution The execution strategy is simulated in the digital twin. After verifying that the N-1 power supply safety criteria are met, instructions are issued through the GOOSE protocol: disconnect the output circuit breaker of the #3 battery pack and close the standby battery pack interconnection switch to achieve seamless switching of power supply to the #1 unit protection system.

[0064] Step 6: Blockchain Recording and Closed-Loop Optimization The entire process of fault detection, decision-making, and execution data is written into the blockchain, and the accumulated data is then used to fine-tune the heterogeneous graph neural network and multi-agent strategy offline.

[0065] 4. Implementation Results (1) Fault warning is issued 30 minutes before the fault occurs, replacing the traditional "fault-after-alarm" mode; (2) The troubleshooting time has been reduced from 2-3 hours to 45 seconds; (3) The core protection system achieves zero-interruption power supply; (4) The workload of manual inspection is reduced by 80%.

[0066] Example 2 In another preferred embodiment, based on Embodiment 1, this embodiment provides a fault detection and troubleshooting system for a hydropower station DC system based on heterogeneous graphs and multi-agents. This system is used in conjunction with the fault detection and troubleshooting method for a hydropower station DC system based on heterogeneous graphs and multi-agents described in Embodiment 1. Figure 2 The system architecture shown is constructed as follows: 1. Application Scenarios This embodiment is applied to a pumped storage power station, whose DC system needs to be adapted to power generation / pumping dual-mode switching, including 6 sets of batteries (2 sets as backup), 12 bidirectional charging devices, and 89 DC feeder circuits.

[0067] 2. System architecture (e.g.) Figure 2 (As shown) This system includes a sensor network module, a multi-source heterogeneous construction module, a digital twin construction module, a heterogeneous graph neural network module, a strategy module, and a blockchain module. These modules are interconnected. (1) Sensor network module Deploy multiple types of sensors: Configure voltage / internal resistance monitoring units for individual battery cells; collect charging and discharging current and ripple from the charging device; install leakage current sensors on DC feeder branches; deploy acceleration sensors (to meet vibration monitoring requirements during operating condition changes) on battery racks and charging device cabinets.

[0068] (2) Multi-source heterogeneous graph construction module After receiving sensor network data, the vibration signal is processed by ensemble empirical mode decomposition (EEMD) to extract characteristic mode functions related to the switching of operating conditions (such as identifying the abnormal vibration energy in the 125Hz frequency band of the #2 charging device); at the same time, it is associated with equipment ledgers and topological relationships to form structured data of "equipment-relationship-attribute".

[0069] (3) Digital Twin Construction Module It includes a digital twin module and a two-way real-time synchronization interface (synchronization frequency 1 time / second) to achieve state synchronization between the twin and the physical system; and dynamically adjusts the path weights of heterogeneous graph elements according to the operating conditions: the power generation condition focuses on the "battery-discharge -> load" path, and the pumping condition focuses on the "charger-charging -> battery" path.

[0070] (4) Heterogeneous neural network module Data input module: Receives structured data from multi-source heterogeneous construction modules; Fault detection module: Learns node and semantic features through heterogeneous graph attention network (HAN), and locates the mechanical connection loosening fault of charging device #2 based on node-level attention coefficient formula (same as in embodiment 1).

[0071] (5) Strategy Module Multi-agent reinforcement learning model: Two sets of MADDPG policy networks are trained. Under power generation conditions, priority is given to ensuring the power supply for the generating unit control system; under pumping conditions, priority is given to ensuring the power supply for the pumps. The loss function of the Critic network in the algorithm is: (8); In the formula, This corresponds to 6 partitioned agents; m corresponds to 6 partitioned agents. For the joint observation of "charger current + battery SOC" under dual operating conditions, This refers to combined actions such as "charger switching and load adjustment"; Digital twin strategy verification module: Simulates the operating condition switching process in the twin to verify the voltage stability constraints of the strategy.

[0072] (6) Blockchain module Record system lifecycle data (operating condition switching data, fault handling logs, etc.) to provide an immutable basis for equipment status assessment.

[0073] 3. Implementation Results and Economic Benefits (1) Successfully issued an early warning of mechanical failure of the #2 charging device, avoiding shutdown caused by disconnection of the connector; (2) When switching operating conditions, the DC voltage fluctuation is reduced from ±12% to ±5%, and the power supply reliability reaches 99.99%; (3) Avoiding unplanned unit shutdowns once saves approximately RMB 1.8 million in costs; (4) The battery replacement cycle is extended from 6 years to 8 years, saving about RMB 600,000 in maintenance costs annually.

[0074] In the preferred embodiment, the multi-source heterogeneous data mentioned in step 1 includes the voltage and internal resistance of individual battery cells, the output ripple of the charging module, the leakage current of each DC feeder branch, the temperature of key nodes, the vibration signal of the switchgear, and the original data of the insulation monitoring device. The above settings comprehensively cover the operating parameters of the DC system, eliminate the blind spots of a single data source, and provide complete information support for fault feature extraction.

[0075] By integrating data from devices such as batteries and feeders, it supports cross-dimensional correlation analysis, accurately identifies coupled fault modes, reduces the risk of misjudgment due to missing data, and improves the comprehensiveness of diagnosis.

[0076] In the preferred embodiment, the edge-side fusion processing in step 1 includes data preprocessing and feature fusion. The data preprocessing specifically involves: using ensemble empirical mode decomposition to process the switchgear vibration signal to extract intrinsic mode functions; and using adaptive Kalman filtering to smooth the battery internal resistance measurement value. These settings accurately extract the intrinsic mode functions of the vibration signal and effectively separate noise and fault characteristic components. Through nonlinear and non-stationary signal decomposition technology, early identification of weak vibration anomalies is achieved, avoiding the analytical errors of traditional methods for complex signals and improving the sensitivity of equipment status monitoring.

[0077] In the preferred embodiment, the feature fusion specifically involves: associating preprocessed data with equipment ledgers, topological relationships, and historical maintenance records based on a knowledge graph to construct a triplet of node-relationship-attribute, forming a unified feature representation; dynamically adjusting filtering parameters to eliminate measurement noise interference and improve the continuity and stability of internal resistance data; and providing a reliable basis for judging internal resistance anomaly thresholds by real-time correction of system noise statistical characteristics, reducing the number of false alarms and enhancing the accuracy of status assessment.

[0078] In the preferred embodiment, the meta-path in step 2.3 includes "battery - (power supply) -> DC feeder - (power supply) -> protection device". By aggregating information through meta-paths, the perception of the risk of power failure of the protection device is enhanced. The above settings quantify the energy transfer and protection dependency between devices, and strengthen the early warning capability of the risk of power failure of the protection device. By aggregating path information, potential fault propagation paths are revealed, and early prediction of the risk of power failure of the protection device is achieved, providing a basis for preventive maintenance.

[0079] In the preferred embodiment, in step 4.3, each agent has an Actor network and a Critic network. The Actor network selects actions based on its own observations, and the Critic network evaluates the value of joint observations and joint actions of all agents. The Critic network updates by minimizing the temporal difference error. With this setup, the Actor network independently selects actions based on local observations, and the Critic network evaluates the value of joint actions through global information. By using the temporal difference error minimization update strategy, the global optimality of the multi-agent collaborative strategy is ensured, improving the system's adaptive capability and decision-making efficiency under complex conditions.

[0080] In the preferred embodiment, the sensor network module is used to collect multi-source operating data of the hydropower station's DC system, and transmit the multi-source operating data to the heterogeneous graph construction module for storage and structured processing to form a multi-source heterogeneous data structure. This configuration, by comprehensively collecting data from key equipment such as batteries and charging modules, ensures information integrity and provides a rich data foundation for subsequent analysis. The structured processing of the heterogeneous data structure eliminates data format differences, improves processing efficiency, supports comprehensive diagnosis across devices and systems, and reduces the risk of misjudgment due to missing data.

[0081] In the preferred embodiment, the digital twin construction module achieves bidirectional data synchronization of the operational status between the constructed digital twin and the physical entity of the hydropower station's DC system through a bidirectional real-time synchronization interface. This configuration, by mapping the operational status of the physical system in real time, constructs a high-fidelity virtual model, supporting the pre-verification of strategies in the virtual environment. The bidirectional synchronization interface ensures that the virtual model and the physical entity are consistent, avoiding decision-making biases caused by model lag, improving the accuracy of strategy feasibility assessment, and reducing the operational risks of the actual system.

[0082] In a preferred embodiment, the data input module is used to input the structured multi-source heterogeneous data output by the multi-source heterogeneous graph construction module to the heterogeneous neural network module to support the heterogeneous neural network module in performing fault detection and localization operations. This configuration, by inputting structured multi-source heterogeneous data to the heterogeneous neural network module, provides multi-dimensional feature support for fault detection. By fusing data features from batteries, feeders, etc., the limitations of a single data source are overcome, improving the ability to identify coupled faults, increasing localization accuracy to the branch level, and reducing the scope and time of investigation.

[0083] In a preferred embodiment, the multi-agent reinforcement learning model is communicatively connected to the heterogeneous neural network module, and is used to generate a corresponding collaborative self-healing strategy based on the fault detection and localization results output by the heterogeneous neural network module. With this configuration, based on the fault results output by the heterogeneous neural network, each agent makes independent decisions through the Actor network, and the Critic network evaluates the value of the joint action, achieving distributed collaboration. Through a global information optimization strategy, local optima are avoided, ensuring the global optimality of the self-healing scheme, shortening fault recovery time, and improving the system's adaptive capability.

[0084] In a preferred embodiment, the digital twin strategy verification module is connected to both the multi-agent reinforcement learning model and the digital twin construction module. This module simulates the execution of a collaborative self-healing strategy within the digital twin and verifies its feasibility. This setup simulates the execution of the collaborative self-healing strategy in a virtual environment, monitors the strategy's impact on the virtual system in real time, and verifies its physical feasibility. By comparing the state changes before and after strategy execution, potential problems can be identified and optimized in advance, avoiding operational risks in the actual system, improving decision-making reliability, and reducing trial-and-error costs.

[0085] In summary, this invention proposes a method and system for fault detection and elimination in hydropower station DC systems based on heterogeneous graphs and multi-agents. It effectively solves the three core problems in the field of fault detection in hydropower station DC systems: incomplete fault feature perception, weak modeling capability of heterogeneous equipment relationships, and low efficiency of multi-fault collaborative decision-making. At the same time, it overcomes the limitations of existing technologies such as single detection dimension, low positioning accuracy, and lack of intelligent decision-making.

[0086] This invention pioneers the application of heterogeneous graph neural networks to fault detection in DC systems of hydropower stations, constructing a heterogeneous graph digital twin that can accurately distinguish different types of equipment and their complex connections, effectively perceiving multi-type fault characteristics. A multi-agent reinforcement learning algorithm is employed to handle collaborative fault decision-making problems, dividing the system into multiple agents and achieving multi-zone power supply collaboration and mutual assistance through Markov game processes, overcoming the limitations of single-agent reinforcement learning. A strategy simulation and security verification method based on digital twin technology is proposed, simulating the execution of collaborative self-healing strategies and performing security verification within the digital twin to ensure the safe execution of control commands; this method is the first of its kind in this field. Edge computing nodes are used to preprocess multi-source heterogeneous data, employing ensemble empirical mode decomposition and adaptive Kalman filtering methods, combined with knowledge graphs for feature fusion, forming a unified feature representation. Based on blockchain's immutable data storage and offline fine-tuning mechanism, continuous optimization of the model and strategy is achieved, adaptable to scenarios such as equipment aging and changing operating conditions, providing a new solution for long-term stable system operation.

[0087] This invention constructs a comprehensive technical framework for fault detection and elimination in hydropower station DC systems. It organically combines heterogeneous graph neural networks, multi-agent reinforcement learning, digital twin technology, edge computing, and blockchain technology to form a complete end-to-end intelligent solution, addressing multiple issues. In fault perception, heterogeneous graph neural networks are used to finely model the differences and complex relationships between heterogeneous equipment. Combined with multi-source heterogeneous data fusion, coupled faults and latent faults are accurately located, overcoming the limitations of traditional detection methods and providing new insights for accurate fault diagnosis. In decision generation, multi-agent reinforcement learning algorithms enable collaborative assistance among power supply zones, generating globally optimal self-healing strategies, overcoming the shortcomings of traditional methods, and improving fault handling efficiency and reliability. In strategy execution, a strategy simulation and security verification method based on digital twins provides dual protection for the safe execution of control commands, preventing malfunctions from escalating and ensuring the safe and stable operation of the system. This security mechanism has significant value in the field of power system fault handling. Through data preprocessing and feature fusion using edge computing nodes, and a blockchain-based model and strategy continuous optimization mechanism, the comprehensiveness and accuracy of fault detection are improved, enabling the system to adapt to various changes during long-term operation.

Claims

1. A method for fault detection and troubleshooting of DC systems in hydropower stations based on heterogeneous graphs and multi-agent systems, characterized in that, Includes the following steps: Step 1: Collect multi-source heterogeneous data from the hydropower station's DC system and perform edge-side fusion processing; Step 2: Construct a heterogeneous digital twin of the DC system of the hydropower station; Step 3: Fault detection and localization based on heterogeneous graph neural networks; Step 4: Generate a collaborative self-healing strategy based on multi-agent reinforcement learning; Step 5: Verify and securely execute the collaborative self-healing strategy based on the digital twin.

2. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that: The multi-source heterogeneous data mentioned in step 1 includes the individual battery cell voltage and internal resistance, charging module output ripple, leakage current of each DC feeder branch, critical node temperature, switchgear vibration signal, and raw data from insulation monitoring devices.

3. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that: The edge-side fusion process in step 1 includes data preprocessing and feature fusion. Specifically, the data preprocessing involves: using ensemble empirical mode decomposition to process the switchgear vibration signal to extract intrinsic mode functions; and using adaptive Kalman filtering to smooth the battery internal resistance measurement value.

4. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 3, characterized in that: The feature fusion specifically involves associating preprocessed data with equipment ledgers, topological relationships, and historical maintenance records based on a knowledge graph, constructing a triplet of node-relationship-attribute to form a unified feature representation.

5. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that, The construction of the heterogeneous graph digital twin in step 2 includes: Step 2.1: Define the node types of the heterogeneous graph. Node types include batteries, chargers, AC power supplies, DC feeders, and loads. Step 2.2: Define the edge relationship types of the heterogeneous graph. Edge relationship types include electrical connection, control relationship, physical proximity, and logical dependency. Step 2.3: Enrich node features based on meta-path semantics.

6. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 5, characterized in that, The meta-path in step 2.3 includes "battery - (power supply) -> DC feeder - (power supply) -> protection device". Information is aggregated through the meta-path to enhance the perception of the risk of power failure of the protection device.

7. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that, The heterogeneous graph neural network mentioned in step 3 is a heterogeneous graph attention network model, and the fault detection and localization include: Step 3.1: Decompose the heterogeneous graph into multiple isogeneous subgraphs according to the preset meta-path; Step 3.2: Perform node-level attention learning on each isomorphic subgraph and calculate the attention distribution of a node on its neighbors; Step 3.3: Perform semantic-level attention learning, fuse node embeddings from different meta-paths, and obtain the final node representation; Step 3.4: Input the node representation into the classifier to predict the fault type and fault probability of each node.

8. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 7, characterized in that, In step 3.2, when calculating the attention distribution of a node on its neighbors, for the metapath... ,node and nodes Attention coefficient The calculation is as follows: （1）； In the formula, , , These are nodes , , initial characteristics, Metapath A specific weight matrix, It is an attention vector. It is the Leaky ReLU activation function. It is a node In metapath The set of neighbors below.

9. The method for fault detection and troubleshooting of a hydropower station DC system based on heterogeneous graphs and multi-agent systems according to claim 7, characterized in that, The final node representation described in step 3.3 for: （2）； In the formula, Metapath The weights; Metapath At the node The representation of; It is a set of metapaths.

10. The method for fault detection and troubleshooting of a hydropower station DC system based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that, Step 4, which generates a collaborative self-healing strategy based on multi-agent reinforcement learning, includes: Step 4.1: Divide the DC system into multiple intelligent agents according to the power supply zone. The intelligent agents include the generator DC section Agent, the common section Agent, and the backup battery group Agent. Step 4.2: Define a Markov game process for multi-agent cooperation. The state space of the Markov game process is the joint local observations perceived by all agents. The action space includes load shedding, requesting support, isolating faults, and switching to backup power. The reward function is... Including global rewards Local rewards and penalty items , means as follows: （3）； In the formula, express The current state of the system (such as the current operating status of the DC system, load distribution, etc.); express Actions performed by the system at all times (such as load shedding, power supply adjustment, etc.); Global rewards Weighting coefficients; For local rewards Weighting coefficients; For penalty items Weighting coefficients; Step 4.3: Train the system using the multi-agent deep deterministic policy gradient algorithm; Step 4.4: Each agent shares intent through a communication mechanism based on the distributed execution strategy observed in real time.

11. The method for fault detection and troubleshooting of a hydropower station DC system based on heterogeneous graphs and multi-agent systems according to claim 10, characterized in that, In step 4.3, each agent has an Actor network and a Critic network. The Actor network selects actions based on its own observations, and the Critic network evaluates the value of the joint observations and joint actions of all agents. The Critic network is updated by minimizing the temporal difference error, as shown below: （4）； in, （5）； In the formula, Indicates the first The loss function of an agent-based Critic network; For the first Trainable parameters of an agent Critic network; For the mathematical expectation operator, the average is calculated based on samples in the empirical replay buffer D; For the first The current evaluation value of the Critic network for each agent; For joint observation by all agents; Indicates all The joint action of several intelligent agents; For timely rewards following the execution of joint operations; For the next joint observation after the joint action is executed; The target value for time-series difference; For the first Immediate rewards for actions performed jointly by agents in a Critic network; Discount factor; Indicates the first The target network evaluation value of the Critic network for each agent; This represents the joint action of all agents in the next state; Indicates the next state The joint action of several intelligent agents; Indicates the first The Actor target network policy function for each agent; For all Local observation of an agent.

12. The method for fault detection and troubleshooting of hydropower station DC systems based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that, Step 5, which verifies and securely executes the collaborative self-healing strategy based on a digital twin, includes: Step 5.1: Simulate the execution of the collaborative self-healing strategy in the digital twin to predict the dynamic response of the system; Step 5.2: Perform a security check to verify whether the collaborative self-healing strategy meets the N-1 power supply safety criteria and the hard constraint of uninterrupted core protection power supply. If the check fails, the strategy rollback will be triggered and a conservative predefined strategy will be adopted. Step 5.3: After successful verification, control commands are sent to the intelligent circuit breaker and contactor devices via the GOOSE protocol to complete fault isolation and network reconfiguration.

13. The method for fault detection and troubleshooting of a hydropower station DC system based on heterogeneous graphs and multi-agent systems according to claim 1, characterized in that, It also includes step 6, which involves constructing a closed-loop optimization mechanism, specifically: Step 6.1: Record the entire fault handling process data to the blockchain. The entire process data includes detection data, decision data, and execution result data. Step 6.2: Utilize the accumulated field data to periodically fine-tune the heterogeneous graph neural network model and the multi-agent reinforcement learning strategy offline.

14. A fault detection and troubleshooting system for a hydropower station DC system based on heterogeneous graphs and multi-agents, which is used to implement the fault detection and troubleshooting method for a hydropower station DC system based on heterogeneous graphs and multi-agents as described in any one of claims 1 to 13, characterized in that: It includes a multi-source heterogeneous graph construction module, a digital twin construction module, a heterogeneous neural network module, a policy module, and a blockchain module. The modules are interconnected for data transmission and sharing. The multi-source heterogeneous graph construction module includes a sensor network module and a heterogeneous graph construction module. The digital twin construction module includes a digital twin module and a bidirectional real-time synchronization interface. The heterogeneous neural network module includes a data input module and a fault detection and localization module. The policy module includes a multi-agent reinforcement learning model and a digital twin policy verification module.

15. The hydropower station DC system fault detection and troubleshooting system based on heterogeneous graphs and multi-agents as described in claim 14, characterized in that: The sensor network module is used to collect multi-source operation data of the hydropower station's DC system, transmit the multi-source operation data to the heterogeneous graph construction module for storage, and perform structured processing to form a multi-source heterogeneous data structure.

16. The hydropower station DC system fault detection and troubleshooting system based on heterogeneous graphs and multi-agent systems according to claim 14, characterized in that: The digital twin construction module achieves bidirectional data synchronization of the operational status between the constructed digital twin and the physical entity of the hydropower station's DC system through a bidirectional real-time synchronization interface.

17. The hydropower station DC system fault detection and troubleshooting system based on heterogeneous graphs and multi-agent systems according to claim 14, characterized in that: The data input module is used to input the structured multi-source heterogeneous data output by the multi-source heterogeneous graph construction module into the heterogeneous neural network module to support the heterogeneous neural network module in performing fault detection and localization operations.

18. The hydropower station DC system fault detection and troubleshooting system based on heterogeneous graphs and multi-agents as described in claim 14, characterized in that: The multi-agent reinforcement learning model is communicatively connected to the heterogeneous neural network module and is used to generate corresponding collaborative self-healing strategies based on the fault detection and localization results output by the heterogeneous neural network module.

19. The hydropower station DC system fault detection and troubleshooting system based on heterogeneous graphs and multi-agent systems according to claim 18, characterized in that: The digital twin strategy verification module is connected to the multi-agent reinforcement learning model and the digital twin construction module, respectively, and is used to simulate the execution of a collaborative self-healing strategy in the digital twin and verify the feasibility of the strategy.

Images