Intelligent ground fault diagnosis method and device for photovoltaic power station

Abstract

The invention discloses an intelligent ground fault diagnosis method and device for a photovoltaic power station. According to the method, a static graph structure of a power station is constructed by collecting equipment state and environment data in real time, and node initial feature vectors are generated by using a graph encoder; the core is that a pre-trained state reasoning network is adopted to carry out K-step iterative reasoning, the node state is dynamically updated in each step, the causal edge weight between nodes is recalculated, and the propagation path of a fault in a power station network is accurately simulated. And finally, realizing accurate prediction and positioning of the fault based on the output node state vector and the dynamic causal edge weight matrix. According to the method, the limitation that a traditional method depends on artificial experience is broken through, the complex causal relationship in the system is automatically captured through the dynamic graph model, high-precision and automatic intelligent fault early warning is achieved, and the operation reliability and the operation and maintenance efficiency of the photovoltaic power station are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power grid equipment operation monitoring technology, and in particular to an intelligent grounding fault diagnosis method and device for photovoltaic power plants. Background Technology

[0002] As a crucial component of clean energy, photovoltaic (PV) power plants are increasingly large-scale and complex. A typical PV power plant consists of a large number of PV strings, combiner boxes, inverters, transformers, and other electrical equipment connected by intricate physical and electrical links to form a cohesive system. Grounding faults are among the most common and extremely dangerous types of faults in PV power plant operation. They can lead to equipment damage, reduced power generation efficiency, and even fires or personal injury accidents. Therefore, rapid and accurate grounding fault diagnosis is essential for ensuring the safe and stable operation of PV power plants and reducing maintenance costs.

[0003] Currently, the methods for diagnosing grounding faults in photovoltaic power plants can be mainly classified into the following categories: Traditional threshold-based methods: This is currently the most mainstream technology in industrial applications. Its core idea is to monitor the electrical parameters (such as insulation resistance to ground, leakage current, etc.) of key equipment (such as inverters). When the monitored value exceeds a preset safety threshold, a fault alarm is triggered. However, this method has significant limitations: High false alarm rate: Environmental factors (such as humidity, dust) can easily cause normal fluctuations in parameters such as insulation resistance, which can easily be misjudged as faults. Poor location capability: This method can usually only determine that the fault occurs in a general area (such as a branch circuit), and cannot accurately locate the specific faulty equipment (such as which combiner box or string), let alone reveal the propagation path and root cause of the fault. Isolated analysis: It treats each monitoring point as an independent entity, ignoring the fact that a photovoltaic power station is a tightly coupled system with complex electrical connections and fault propagation effects between equipment. A fault in one device may trigger a chain reaction of abnormal parameters in neighboring devices, and traditional methods cannot capture this systemic interaction.

[0004] Data-driven diagnostic models: With the development of artificial intelligence, some studies have attempted to use machine learning models (such as support vector machines and neural networks) to analyze historical operational data to diagnose faults. While these methods have improved identification capabilities to some extent, they still have inherent drawbacks: "Black box" models: Their decision-making process is opaque, making it difficult for maintenance personnel to understand why the model makes a certain diagnostic conclusion, lacking credibility, and hindering subsequent fault investigation and maintenance decisions. Ignoring topology: Traditional machine learning models treat input data as independent and identically distributed vectors, completely ignoring the inherent physical connections and electrical logic relationships between devices within a photovoltaic power station. The model cannot understand how faults propagate through cables, communication lines, etc., in the actual topology, resulting in a lack of physical interpretability in diagnostic results and difficulties in root cause analysis.

[0005] Preliminary Exploration Based on Graph Neural Networks: Graph Neural Networks (GNNs) have shown potential in the diagnosis of complex systems due to their powerful relational modeling capabilities. Existing technologies have attempted to model power plants as graph structures for fault detection. However, most existing methods use static graph structures, where the edges (representing relationships between devices) are fixed, and weights are typically pre-set based on physical distance or fixed connections. This does not match the dynamic characteristics of real fault propagation: Lack of causal dynamism: In actual operation, the impact intensity and propagation path of a fault are not constant; they change dynamically with the system's operating state and environmental conditions. Static graphs cannot represent this dynamic, causal relationship. Insufficient information aggregation capabilities: Traditional GNN message passing mechanisms (such as first-order neighbor aggregation) may not be able to fully capture the remote influence of multi-hop neighbors (i.e., indirectly connected devices), limiting their effectiveness in locating remote fault sources in large photovoltaic power plants. Single state update mechanism: When updating node states, the failure to effectively distinguish and integrate new information from neighbors with the node's own state maintenance needs may result in an unrobust state representation that is easily affected by noisy data.

[0006] Therefore, there is an urgent need in this field for a new method that can deeply integrate photovoltaic power plant topology knowledge, dynamically capture fault causal propagation relationships, and provide high-precision, interpretable fault diagnosis results to overcome the above-mentioned deficiencies of existing technologies. Summary of the Invention

[0007] This invention provides an intelligent grounding fault diagnosis method and device for photovoltaic power plants, which solves the problems of existing technologies that rely on human experience for diagnosis and are difficult to locate the root cause of faults in real time and accurately. It realizes automated and high-precision fault diagnosis and early warning based on dynamic causal graph reasoning.

[0008] In a first aspect, the present invention provides a method for intelligent grounding fault diagnosis in photovoltaic power plants, the method comprising: Acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; Based on the physical connection relationships and electrical logic relationships of the photovoltaic power station, a static diagram structure is determined; wherein, the nodes in the static diagram structure represent the various devices in the photovoltaic power station; The real-time running data is transformed using a graph structure encoder to obtain the initial feature vectors corresponding to each node; wherein, the initial feature vectors include: a device state embedding vector representing the device state and an environmental state embedding vector representing the environmental conditions; The initial feature vector and the static graph structure are input into a pre-trained state inference network, and K-step iterative inference is performed to obtain the final node state vector and the final dynamic causal edge weight matrix between nodes. Specifically, the state inference network outputs the final node state vector and the final dynamic causal edge weight matrix between nodes through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vector and the final dynamic causal edge weight matrix between nodes. Based on the final node state vector and the final dynamic causal edge weight matrix between each node, the fault prediction estimate is obtained.

[0009] In conjunction with the first aspect, in one possible implementation, determining the static diagram structure based on the physical connection relationships and electrical logic relationships of the photovoltaic power station includes: The physical connection relationships include: physical wiring and communication connections between devices; the electrical logic relationships include: current paths, voltage associations, or power flow directions. Based on the physical connection relationship, at least two of the inverters, combiner boxes, photovoltaic strings, transformers and circuit breakers in the photovoltaic power station are constructed as nodes in the static graph structure, and the actual relationship between them is constructed as edges in the static graph structure.

[0010] In conjunction with the first aspect, in one possible implementation, the step of using a graph structure encoder to transform the real-time running data to obtain the initial feature vectors corresponding to each node includes: The graph structure encoder includes: a device status coding branch and an environment status coding branch; The equipment state coding branch extracts features from at least one of the current, voltage, power, temperature, and insulation resistance values ​​of the equipment in the photovoltaic power station to generate an equipment state embedding vector; The environmental state coding branch extracts features from at least one of the following: ambient light intensity, ambient temperature, humidity, and dust concentration, to generate an environmental state embedding vector. The device state embedding vector and the environment state embedding vector corresponding to the same node are concatenated to form the initial feature vector of that node.

[0011] In conjunction with the first aspect, in one possible implementation, the state reasoning network includes a multi-hop perception aggregation unit, a bidirectional information fusion unit, and a causal structure learning unit connected in sequence. The multi-hop perception aggregation unit includes a hierarchical attention calculation module, a hop count fusion gating module, and a residual message generation module connected in sequence. The hierarchical attention calculation module is used to calculate an independent attention weight vector for each hop neighbor of the central node; The hop count fusion gating module is used to generate and fuse the gating vector of attention weights for each hop; The residual message generation module is used to perform a weighted summation of the hidden states of multi-hop neighbors using the gating vector, and perform a residual connection between the result and the aggregation result of the first-order neighbors to generate the final message vector. The bidirectional information fusion unit includes, in sequence, a forward update unit, a reverse maintenance unit, and a state fusion unit; The forward update unit is used to calculate the state update amount based on the node's current state vector and message vector. The reverse maintenance unit is used to calculate the state maintenance amount based on the difference between the current state vector and the message vector of the node; the reverse maintenance unit is a parameterized network independent of the forward update unit. The state fusion unit is used to weight and fuse the state update quantity and the state maintenance quantity to output the updated state vector of the node. The causal structure learning unit includes, in sequence, a causal contribution evaluation unit, a static prior injection unit, and a dynamic sparsity processing unit. The causal contribution assessment unit is used to calculate the causal influence score between nodes based on the difference between the state vectors corresponding to the two nodes. The static prior injection unit is used to take the original connection weights of the static graph structure as prior knowledge and concatenate them with the causal influence score to obtain the concatenated feature vector. The dynamic sparsity processing unit is used to normalize the concatenated feature vectors and apply a threshold-based function for sparsification, ultimately outputting a dynamic causal edge weight matrix.

[0012] In conjunction with the first aspect, in one possible implementation, the step of generating a message vector for each node by weighted summing of the state vectors of its neighboring nodes based on the dynamic causal edge weight matrix of the current iteration step includes: The following steps are performed by the multi-hop sensing aggregation unit: For the central node and the central node at Neighbor nodes within the jump , ,implement: Using the hierarchical attention calculation module, an independent learnable weight matrix is ​​set for each hop. and attention vector And calculate the unstandardized attention coefficient. ; The attention coefficient for each jump Normalization is performed to obtain the normalized attention weights for each hop. ; Based on normalized attention weights and the learnable weight matrix Neighbor features after linear transformation Perform a weighted summation operation to obtain the preliminary aggregated message for the current jump. ; Using the hop count fusion gating module, based on the normalized attention weights of each hop... The average attention intensity of each jump was calculated. ; The average of each jump Concatenate into an average attention intensity vector Furthermore, a K-dimensional gated vector is generated through fully connected layer transformation and the Sigmoid activation function. ; Based on the preliminary aggregated messages from each hop and gate vector Perform a weighted fusion operation to obtain the central node. Final message vector .

[0013] In conjunction with the first aspect, in one possible implementation, the step of fusing the current state vector of each node with its corresponding message vector using an update function to obtain the updated state vector of that node includes: For each node that serves as the central node , will node Current state vector With the message vector generated by the multi-hop sensing aggregation unit As input, the following steps are performed by the bidirectional information fusion unit: Using the forward update unit, the node Current state vector and message vector Perform concatenation and linear transformation, and compute nodes using the Sigmoid activation function. Forward update ; Using the reverse maintenance unit, according to the node Current state vector and message vector The difference calculation is performed to obtain the node. State maintenance quantity ; Using the state fusion unit, based on the node State maintenance quantity For the current state vector Modulation is performed, and the modulation result is compared with the node. message vector The candidate states are obtained by fusion. Forward update With state maintenance quantity The data is stitched together, and dynamic fusion weights are calculated using a lightweight perceptron. ; Based on dynamic fusion weights For forward update volume Modulated candidate states and state-maintaining quantities The current modulated state is weighted and fused to output the node. The updated state vector .

[0014] In conjunction with the first aspect, in one possible implementation, the step of recalculating the weights of the dynamic causal edges between nodes based on the updated state vectors of all nodes, updating the dynamic causal edge weight matrix, and obtaining the final node state vectors and the final dynamic causal edge weight matrix between each node includes: The following steps are performed through the causal structure learning unit: For any two nodes connected by a dynamic edge and nodes ,implement: Using the aforementioned causal contribution evaluation unit, the nodes are obtained respectively. and nodes The node updated in the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector ; Based on the node updated in the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector Calculate the nodes respectively and nodes The corresponding first change Second change And use a multilayer perceptron to analyze the first change. and the second change Perform nonlinear feature mapping to obtain higher-order feature representations of the first and second transform quantities; The higher-order feature representations of the first and second transform quantities are concatenated, and the nodes are calculated based on the concatenation result. For nodes causal effect score ; Using the static prior injection unit, nodes in the static graph structure are obtained. and nodes Corresponding initial weights The initial weights With the causal effect score Perform linear weighted fusion to obtain the fusion score. ; Using the dynamic sparsity processing unit, the fusion score is... Normalize and apply a learnable threshold. The sparsification function will be below the threshold. The fusion score is set to zero, resulting in the node. and New weights of dynamic causal edges To update the dynamic causal edge weight matrix.

[0015] In conjunction with the first aspect, in one possible implementation, the fault prediction estimation includes at least one of the following: the fault occurrence confidence of the specified equipment, the equipment health score, and the potential fault root cause nodes and impact paths obtained by backtracking analysis based on the final dynamic causal edge weight matrix.

[0016] In conjunction with the first aspect, in one possible implementation, the training process of the state reasoning network includes: Collect historical operating data and corresponding historical fault tag data of photovoltaic power plants; Construct a historical graph structure corresponding to the historical operational data; Using the historical graph structure and historical operation data as model inputs, and the historical fault label data as supervision targets, the state reasoning network, which includes the multi-hop perception aggregation unit, bidirectional information fusion unit, and causal structure learning unit, is trained end-to-end to minimize the loss function between the predicted results and the true labels.

[0017] Secondly, the present invention provides an intelligent grounding fault diagnosis device for a photovoltaic power station, the device comprising: The data acquisition module is used to acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; The graph structure confirmation module is used to determine the static graph structure based on the physical connection relationship and electrical logic relationship of the photovoltaic power station; wherein, the nodes in the static graph structure represent the equipment in the photovoltaic power station; The data conversion module is used to convert the real-time running data using a graph structure encoder to obtain the initial feature vectors corresponding to each node; wherein, the initial feature vectors include: a device state embedding vector representing the device state and an environmental state embedding vector representing the environmental conditions; The iterative inference module is used to input the initial feature vector and the static graph structure into the pre-trained state inference network, perform K-step iterative inference, and obtain the final node state vector and the final dynamic causal edge weight matrix between each node. The state inference network outputs the final node state vector and the final dynamic causal edge weight matrix between each node through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vector and the final dynamic causal edge weight matrix between each node. The fault prediction module is used to obtain fault prediction estimates based on the final node state vector and the final dynamic causal edge weight matrix between each node.

[0018] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention acquires real-time operational data from photovoltaic power plants, including at least equipment status data representing equipment status and environmental status data representing environmental conditions. By monitoring equipment status and environmental conditions in real time, it can promptly capture abnormal changes, providing the latest and most comprehensive data foundation for fault prediction, thereby enhancing the timeliness and accuracy of predictions and avoiding missed or false alarms due to data delays. Based on the physical connections and electrical logic relationships of the photovoltaic power plant, a static graph structure is determined, where nodes represent various devices within the photovoltaic power plant. Constructing a graph structure based on actual physical and electrical relationships accurately reflects the topological layout of the power plant, ensuring that subsequent analysis conforms to the real system structure, improving the accuracy and interpretability of fault location, and reducing erroneous reasoning caused by structural misunderstandings. A graph structure encoder is used to transform the real-time operational data, obtaining initial feature vectors corresponding to each node. These initial feature vectors include equipment status embedding vectors representing equipment status and environmental status embedding vectors representing environmental conditions. Converting the raw data into embedding vectors through the encoder allows for the extraction of high-level feature representations, reducing data dimensionality while retaining key information, facilitating efficient processing by graph neural networks, improving computational efficiency and model generalization ability, and enhancing the capture of complex operating modes. The initial feature vectors and static graph structure are input into a pre-trained state inference network, and K-step iterative inference is performed to obtain the final node state vectors and the final dynamic causal edge weight matrix between nodes. The state inference network outputs the final node state vectors and the final dynamic causal edge weight matrix between nodes through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated. Through the iterative inference process, the model can gradually optimize node states and edge weights, dynamically capture causal relationships between nodes and system state evolution, enhance the identification of fault propagation paths, improve the robustness and adaptability of inference, and thus more accurately reflect the real-time operating status and reduce the risk of misjudgment of faults. Based on the final node state vector and the final dynamic causal edge weight matrix between nodes, a fault prediction estimate is obtained. Using the refined state vector and dynamic edge weights, potential fault points and risk levels can be accurately identified, enabling proactive fault prediction. This helps maintenance personnel take preventative maintenance measures, reducing power plant downtime and improving operational reliability and economic efficiency. Through the organic combination of these steps, the entire method achieves efficient and accurate fault prediction for photovoltaic power plants, enhancing the system's intelligent operation and maintenance level. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the steps of an intelligent grounding fault diagnosis method for photovoltaic power plants, provided in an embodiment of the present invention. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In a first aspect, the present invention provides an intelligent grounding fault diagnosis method for photovoltaic power plants, see [link to relevant documentation]. Figure 1 The method includes the following steps S101 to S105.

[0022] S101, acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; For example, real-time operational data of a photovoltaic power plant is acquired through a sensor network and Supervisory Control and Data Acquisition (SCADA) system deployed in the plant. This real-time operational data includes at least: equipment status data characterizing the equipment's condition and environmental status data characterizing the environmental conditions.

[0023] Equipment status data includes, but is not limited to: DC / AC side current, voltage, output power, internal temperature of the inverter, and the insulation resistance to ground value monitored by the system.

[0024] Environmental condition data includes, but is not limited to: light intensity, ambient temperature, humidity, and dust concentration.

[0025] S102, Determine the static diagram structure based on the physical connection relationship and electrical logic relationship of the photovoltaic power station; wherein, the nodes in the static diagram structure represent the various devices in the photovoltaic power station; Specifically, in step S102, the static diagram structure is determined based on the physical connection relationships and electrical logic relationships of the photovoltaic power station, including: (1) Physical connection relationships include: physical wiring and communication connections between devices; electrical logic relationships include: current paths, voltage associations or power flow directions. (2) Based on the physical connection relationship, at least two of the inverters, combiner boxes, photovoltaic strings, transformers and circuit breakers in the photovoltaic power station are constructed as nodes in the static graph structure, and the actual relationship between them is constructed as edges in the static graph structure.

[0026] For example, a typical static graph structure for a photovoltaic power plant can be constructed as follows: "Photovoltaic strings" are leaf nodes, converging upwards to "combiner box" nodes. Multiple "combiner box" nodes then connect to "inverter" nodes, and the "inverter" nodes ultimately connect to "transformer" nodes. The connecting lines (physical cables and communication lines) between them constitute the edges in the graph. This structure can be represented by an adjacency matrix A_static, whose initial edge weights can be uniformly set to 1, or initialized based on prior knowledge such as the specifications and rated current of the connected cables.

[0027] S103, The graph structure encoder is used to transform the real-time running data to obtain the initial feature vectors corresponding to each node; wherein, the initial feature vectors include: the device state embedding vector representing the device state and the environment state embedding vector representing the environmental conditions. Specifically, in step S103, the real-time running data is transformed using a graph structure encoder to obtain the initial feature vectors corresponding to each node, including the following steps S1031 to S1034.

[0028] S1031, the graph structure encoder includes: a device status coding branch and an environment status coding branch; S1032, the equipment status coding branch extracts features from at least one of the current, voltage, power, temperature and insulation resistance values ​​of the equipment in the photovoltaic power station to generate an equipment status embedding vector; S1033, the environmental state coding branch extracts features from at least one of the environmental light intensity, environmental temperature, humidity, and dust concentration to generate an environmental state embedding vector; S1034, the device state embedding vector and the environment state embedding vector corresponding to the same node are concatenated to form the initial feature vector of the node.

[0029] For example, both the device state encoding branch and the environment state encoding branch can be implemented using a two-layer fully connected neural network. Assume the device state data dimension is... The environmental status data dimension is The expected embedding vector dimension is .but: The device state embedding vector is calculated as follows: , The environment state embedding vector is calculated as follows: , in, , , , The weight matrix is ​​a learnable matrix. , , , This is a learnable bias term.

[0030] Then, the corresponding nodes and The features are concatenated to form the initial feature vector of the node. .

[0031] S104: Input the initial feature vector and static graph structure into the pre-trained state inference network, perform K-step iterative inference, and obtain the final node state vector and the final dynamic causal edge weight matrix between each node.

[0032] Specifically, in step S104, the state reasoning network includes a multi-hop perception aggregation unit, a bidirectional information fusion unit, and a causal structure learning unit connected in sequence. The multi-hop perception aggregation unit includes a hierarchical attention calculation module, a hop count fusion gating module, and a residual message generation module connected in sequence.

[0033] The hierarchical attention calculation module is used to calculate an independent attention weight vector for each hop neighbor of the central node.

[0034] The hop count fusion gating module is used to generate and fuse the gating vectors of attention weights for each hop.

[0035] The residual message generation module is used to perform a weighted summation of the hidden states of multi-hop neighbors using a gating vector, and then perform a residual concatenation of the result with the aggregation result of first-order neighbors to generate the final message vector.

[0036] The bidirectional information fusion unit comprises, in sequence, a forward update unit, a reverse maintenance unit, and a state fusion unit.

[0037] The forward update unit is used to calculate the state update amount based on the node's current state vector and message vector.

[0038] The reverse maintenance unit is used to calculate the state maintenance quantity based on the difference between the current state vector and the message vector of a node; the reverse maintenance unit is a parameterized network independent of the forward update unit.

[0039] The state fusion unit is used to weight and fuse the state update quantity and the state maintenance quantity, and output the updated state vector of the node.

[0040] The causal structure learning unit consists of the following connected units in sequence: causal contribution evaluation unit, static prior injection unit, and dynamic sparsification processing unit.

[0041] The causal contribution assessment unit is used to calculate the causal influence score between nodes based on the difference between the state vectors corresponding to the two nodes.

[0042] The static prior injection unit is used to take the original connection weights of the static graph structure as prior knowledge and concatenate them with the causal influence scores to obtain the concatenated feature vector.

[0043] The dynamic sparsity processing unit is used to normalize the concatenated feature vectors and apply a threshold-based function for sparsification, ultimately outputting a dynamic causal edge weight matrix.

[0044] The state reasoning network performs K-step iterative reasoning, outputting the final node state vector and the final dynamic causal edge weight matrix between nodes. In each iteration, the following operations are performed sequentially: (1) Based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of the neighboring nodes of each node are weighted and summed to generate the message vector of that node; Specifically, based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector, including: The following steps are performed through the multi-hop sensing aggregation unit: (1.1) For the central node and the central node at Neighbor nodes within the jump , ,implement: (1.2) Using the hierarchical attention calculation module, an independent learnable weight matrix is ​​set for each hop. and attention vector And calculate the unstandardized attention coefficient. ; (1.3) Attention coefficient for each jump Normalization is performed to obtain the normalized attention weights for each hop. ; (1.4) Based on normalized attention weights and through the learnable weight matrix Neighbor features after linear transformation Perform a weighted summation operation to obtain the preliminary aggregated message for the current jump. ; (1.5) Using the hop count fusion gating module, based on the normalized attention weights of each hop... The average attention intensity of each jump was calculated. ; (1.6) Mean of each jump Concatenate into an average attention intensity vector Furthermore, a K-dimensional gated vector is generated through fully connected layer transformation and the Sigmoid activation function. ; (1.7) According to the preliminary aggregation messages of each hop and gate vector Perform a weighted fusion operation to obtain the central node. Final message vector .

[0045] (2) Use the update function to fuse the current state vector of each node with its corresponding message vector to obtain the updated state vector of the node; Specifically, the update function is used to fuse the current state vector of each node with its corresponding message vector to obtain the updated state vector of that node, including: For each node that serves as the central node , will node Current state vector With the message vector generated by the multi-hop sensing aggregation unit As input, the following steps are performed by the bidirectional information fusion unit: (2.1) To the update unit, the node Current state vector and message vector Perform concatenation and linear transformation, and compute nodes using the Sigmoid activation function. Forward update ; (2.2) Using the reverse sustaining unit, based on the node Current state vector and message vector The difference calculation is performed to obtain the node. State maintenance quantity ; (2.3) Using the state fusion unit, based on the nodes State maintenance quantity For the current state vector Modulation is performed, and the modulation result is compared with the node. message vector The candidate states are obtained by fusion. Forward update With state maintenance quantity The data is stitched together, and dynamic fusion weights are calculated using a lightweight perceptron. ; (2.4) Based on dynamic fusion weights For forward update volume Modulated candidate states and state-maintaining quantities The current modulated state is weighted and fused to output the node. The updated state vector .

[0046] (3) Based on the updated state vectors of all nodes, recalculate the weights of the dynamic causal edges between nodes, update the dynamic causal edge weight matrix, and obtain the final node state vectors and the final dynamic causal edge weight matrix between each node. Specifically, based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vectors and the final dynamic causal edge weight matrix between each node, including: The following steps are performed using the causal structure learning unit: (3.1) For any two nodes connected by a dynamic edge and nodes ,implement: (3.2) Using the causal contribution assessment unit, obtain the node values ​​respectively. and nodes The node updated in the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector ; (3.3) Update the node based on the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector Calculate the nodes respectively and nodes The corresponding first change Second change And use a multilayer perceptron to analyze the first change. Second change Perform nonlinear feature mapping to obtain higher-order feature representations of the first and second transform quantities; (3.4) Concatenate the higher-order feature representations of the first transform and the second transform, and calculate the node based on the concatenation result. For nodes causal effect score ; (3.5) Obtain nodes in the static graph structure using static prior injection units. and nodes Corresponding initial weights , set initial weight With causal effect score Perform linear weighted fusion to obtain the fusion score. ; (3.6) Using a dynamic sparsity processing unit, the fusion score is processed. Normalize and apply a learnable threshold. The sparsification function will be below the threshold. The fusion score is set to zero, resulting in the node. and New weights of dynamic causal edges To update the dynamic causal edge weight matrix.

[0047] For example, suppose a pre-trained state reasoning network has been deployed on a diagnostic server, with its iteration step K set to 3. The following details a complete diagnostic reasoning process: Initialization: Initialize the feature vectors of all nodes obtained in step S103. This serves as the initial state for each node. Simultaneously, the static graph adjacency matrix A_static determined in step S102 is used as the initial value for the dynamic causal edge weight matrix, i.e. .

[0048] Iterative reasoning (taking the k=1th iteration as an example): a. Message generation (executed by the multi-hop aware aggregation unit): For each central node in the system (e.g., a node numbered...), ... The inverter), this unit will aggregate its Information about the neighboring hops (3 hops in this example).

[0049] Based on the current dynamic causal edge weight matrix The unit first calculates independent attention weights for each hop neighbor using a hierarchical attention calculation module. For example, for a first-order neighbor node... Its non-standardized attention coefficient is calculated as follows: ; in, and These are learnable parameters. This represents vector concatenation. Then, all weights are normalized: .

[0050] Similarly, calculate the normalized attention weights for the second and third hops. and .

[0051] Next, the hop count fusion gating module calculates the average attention intensity for each hop, for example... And generate the gate vector: .

[0052] Finally, the residual message generation module integrates the information from each hop to generate nodes. The final message vector is represented as: , in, It is an aggregated message of first-order neighbors.

[0053] b. State update (executed by the bidirectional information fusion unit): Node Using the received message vector Update your own status.

[0054] The forward update unit calculates the update amount: .

[0055] The reverse sustaining unit (an independent MLP) calculates the sustaining amount: .

[0056] The state fusion unit first calculates the candidate states: ; in, It represents the Hadamardi (or Hadama) stack.

[0057] Subsequently, the dynamic fusion weights are calculated: .

[0058] Finally, output node Updated state vector: .

[0059] c. Causal edge weight update (executed by the causal structure learning unit): After all node states have been updated, this unit recalculates the weights of the dynamic causal edges.

[0060] For any pair of nodes connected by an edge and The causal contribution assessment units calculate their state changes: , .

[0061] Subsequently, a causal effect score is calculated using a multilayer perceptron: .

[0062] The static prior injection unit integrates static priors with causal scores: , in, It is a learnable trade-off parameter.

[0063] The dynamic sparsity processing unit normalizes the fusion score: , And apply a learnable threshold-based approach. Sparsification: .

[0064] All new edge weights This constitutes the dynamic causal edge weight matrix after the first iteration.

[0065] Subsequent iterations and output: The node states obtained in the first iteration and dynamic causal edge weight matrix As input for the second iteration (k=2), steps a, b, and c are repeated. This process continues until the preset K=3 iterations are completed. Finally, the final node state vector is output. and the final dynamic causal edge weight matrix At this point, the state reasoning network has completed the deep reasoning and dynamic modeling of causal relationships for the operating state of the photovoltaic power plant.

[0066] The training process of a state reasoning network includes: (1) Collect historical operating data and corresponding historical fault tag data of photovoltaic power plants; (2) Construct a historical graph structure corresponding to historical operational data; (3) Using historical graph structure and historical operation data as model inputs and historical fault label data as supervision targets, end-to-end training is performed on the state reasoning network containing multi-hop perception aggregation unit, bidirectional information fusion unit and causal structure learning unit to minimize the loss function between the prediction result and the real label.

[0067] For example, collect sufficient historical operating data and corresponding historical fault label data of the photovoltaic power plant. Construct a corresponding historical graph structure. Use the Adam optimizer with cross-entropy loss function as the loss function to train the state inference network end-to-end. The training objective is to minimize the difference between the fault prediction results output by the model and the actual fault labels. The training process can be described as follows: ; in, For model prediction, For real labels, For model parameters, This is the L2 regularization coefficient, used to prevent overfitting.

[0068] S105. Based on the final node state vector and the final dynamic causal edge weight matrix between each node, the fault prediction estimate is obtained.

[0069] Fault prediction estimation includes at least one of the following: the confidence level of failure occurrence of the specified equipment, the equipment health score, and the potential root cause nodes and impact paths obtained by backtracking analysis based on the final dynamic causal edge weight matrix.

[0070] For example, fault prediction estimates can be obtained in the following ways: Equipment failure confidence: The final state vector of each node. Input a Softmax classifier: The output is the confidence level of the device in different health states (such as normal, warning, fault).

[0071] Equipment health score: can be defined as an inverse indicator of fault confidence.

[0072] Potential root cause nodes and impact paths: First, locate the node with the highest confidence in the fault.

[0073] Then, in the final dynamic causal edge weight matrix In this process, starting from the node, backtracking is performed along the edges with higher weights in the incoming edges (edges pointing to the node) to recursively find the upstream node that has the greatest impact on it, until one or more nodes with only high outgoing edge weights and very low incoming edge weights are found. These nodes are the potential root cause nodes of the failure.

[0074] The nodes and edges traversed during the backtracking process constitute the potential failure impact path.

[0075] In a specific simulation embodiment provided by this invention, a simulated photovoltaic power station model containing 200 strings, 20 combiner boxes, 10 inverters, and 1 transformer was built using MATLAB / Simulink. Single-point grounding faults, multi-point grounding faults, and various scenarios under different environmental noise interference were simulated, generating a total of 15,000 sets of labeled sample data.

[0076] The method of this invention (Ours) is compared with traditional thresholding methods, Support Vector Machines (SVM), Graph Convolutional Networks (GCN), and Graph Attention Networks (GAT). Evaluation metrics include fault detection accuracy, F1 score, and root cause localization accuracy.

[0077] The experimental results are shown in the table below:

[0078] Simulation results show that the method of the present invention is significantly better than the comparative methods in terms of the accuracy of fault detection and the precision of root cause localization, verifying the effectiveness of the proposed dynamic causal reasoning and multi-hop perception aggregation mechanism.

[0079] Secondly, the present invention provides an intelligent grounding fault diagnosis device for photovoltaic power plants, the device comprising: The data acquisition module is used to acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; The graph structure confirmation module is used to determine the static graph structure based on the physical connection relationships and electrical logic relationships of the photovoltaic power station; wherein, the nodes in the static graph structure represent the various devices in the photovoltaic power station; The data conversion module is used to convert real-time running data using a graph structure encoder to obtain the initial feature vectors corresponding to each node. The initial feature vectors include: device state embedding vectors representing device status and environmental state embedding vectors representing environmental conditions. The iterative inference module is used to input the initial feature vector and static graph structure into the pre-trained state inference network, perform K-step iterative inference, and obtain the final node state vector and the final dynamic causal edge weight matrix between the nodes. Specifically, the state inference network outputs the final node state vector and the final dynamic causal edge weight matrix between the nodes through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vector and the final dynamic causal edge weight matrix between the nodes. The fault prediction module is used to obtain fault prediction estimates based on the final node state vector and the final dynamic causal edge weight matrix between each node.

[0080] For example, the device can be integrated into a local monitoring server or cloud-based diagnostic platform of a photovoltaic power plant. The data acquisition module reads data from field devices via industrial communication protocols (such as Modbus, IEC 61850); the graph structure verification module automatically generates a static graph structure based on a predefined power plant configuration library; the data conversion module and the iterative inference module are executed after loading a pre-trained model by a deep learning framework (such as PyTorch, TensorFlow) deployed on the server; the fault prediction module presents the diagnostic results (such as a list of faulty equipment, health reports, and causal path diagrams) to the user through a human-machine interface (HMI) or an operation and maintenance management system, and can trigger automatic alarms.

[0081] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for intelligent grounding fault diagnosis in photovoltaic power plants, characterized in that, include: Acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; Based on the physical connection relationships and electrical logic relationships of the photovoltaic power station, a static diagram structure is determined; wherein, the nodes in the static diagram structure represent the various devices in the photovoltaic power station; The real-time running data is transformed using a graph structure encoder to obtain the initial feature vectors corresponding to each node; wherein, the initial feature vectors include: a device state embedding vector representing the device state and an environmental state embedding vector representing the environmental conditions; The initial feature vector and the static graph structure are input into a pre-trained state inference network, and K-step iterative inference is performed to obtain the final node state vector and the final dynamic causal edge weight matrix between nodes. Specifically, the state inference network outputs the final node state vector and the final dynamic causal edge weight matrix between nodes through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vector and the final dynamic causal edge weight matrix between nodes. Based on the final node state vector and the final dynamic causal edge weight matrix between each node, the fault prediction estimate is obtained.

2. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 1, characterized in that, The step of determining the static diagram structure based on the physical connection relationships and electrical logic relationships of the photovoltaic power station includes: The physical connection relationships include: physical wiring and communication connections between devices; the electrical logic relationships include: current paths, voltage associations, or power flow directions. Based on the physical connection relationship, at least two of the inverters, combiner boxes, photovoltaic strings, transformers and circuit breakers in the photovoltaic power station are constructed as nodes in the static graph structure, and the actual relationship between them is constructed as edges in the static graph structure.

3. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 1, characterized in that, The process of using a graph structure encoder to transform the real-time running data to obtain the initial feature vectors corresponding to each node includes: The graph structure encoder includes: a device status coding branch and an environment status coding branch; The equipment state coding branch extracts features from at least one of the current, voltage, power, temperature, and insulation resistance values ​​of the equipment in the photovoltaic power station to generate an equipment state embedding vector; The environmental state coding branch extracts features from at least one of the following: ambient light intensity, ambient temperature, humidity, and dust concentration, to generate an environmental state embedding vector. The device state embedding vector and the environment state embedding vector corresponding to the same node are concatenated to form the initial feature vector of that node.

4. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 1, characterized in that, The state reasoning network includes a multi-hop perception aggregation unit, a bidirectional information fusion unit, and a causal structure learning unit connected in sequence. The multi-hop perception aggregation unit includes a hierarchical attention calculation module, a hop count fusion gating module, and a residual message generation module connected in sequence. The hierarchical attention calculation module is used to calculate an independent attention weight vector for each hop neighbor of the central node; The hop count fusion gating module is used to generate and fuse the gating vector of attention weights for each hop; The residual message generation module is used to perform a weighted summation of the hidden states of multi-hop neighbors using the gating vector, and perform a residual connection between the result and the aggregation result of the first-order neighbors to generate the final message vector. The bidirectional information fusion unit includes, in sequence, a forward update unit, a reverse maintenance unit, and a state fusion unit; The forward update unit is used to calculate the state update amount based on the node's current state vector and message vector. The reverse maintenance unit is used to calculate the state maintenance amount based on the difference between the current state vector and the message vector of the node; the reverse maintenance unit is a parameterized network independent of the forward update unit. The state fusion unit is used to weight and fuse the state update quantity and the state maintenance quantity to output the updated state vector of the node. The causal structure learning unit includes, in sequence, a causal contribution evaluation unit, a static prior injection unit, and a dynamic sparsity processing unit. The causal contribution assessment unit is used to calculate the causal influence score between nodes based on the difference between the state vectors corresponding to the two nodes. The static prior injection unit is used to take the original connection weights of the static graph structure as prior knowledge and concatenate them with the causal influence score to obtain the concatenated feature vector. The dynamic sparsity processing unit is used to normalize the concatenated feature vectors and apply a threshold-based function for sparsification, ultimately outputting a dynamic causal edge weight matrix.

5. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 4, characterized in that, The dynamic causal edge weight matrix based on the current iteration step is used to perform a weighted summation of the state vectors of each node's neighboring nodes to generate the message vector of that node, including: The following steps are performed by the multi-hop sensing aggregation unit: For the central node and the central node at Neighbor nodes within the jump , ,implement: Using the hierarchical attention calculation module, an independent learnable weight matrix is ​​set for each hop. and attention vector And calculate the unstandardized attention coefficient. ; The attention coefficient for each jump Normalization is performed to obtain the normalized attention weights for each hop. ; Based on normalized attention weights and the learnable weight matrix Neighbor features after linear transformation Perform a weighted summation operation to obtain the preliminary aggregated message for the current jump. ; Using the hop count fusion gating module, based on the normalized attention weights of each hop... The average attention intensity of each jump was calculated. ; The average of each jump Concatenate into an average attention intensity vector Furthermore, a K-dimensional gated vector is generated through fully connected layer transformation and the Sigmoid activation function. ; Based on the preliminary aggregated messages from each hop and gate vector Perform a weighted fusion operation to obtain the central node. Final message vector .

6. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 4, characterized in that, The step of fusing the current state vector of each node with its corresponding message vector using an update function to obtain the updated state vector of that node includes: For each node that serves as the central node , will node Current state vector With the message vector generated by the multi-hop sensing aggregation unit As input, the following steps are performed by the bidirectional information fusion unit: Using the forward update unit, the node Current state vector and message vector Perform concatenation and linear transformation, and compute nodes using the Sigmoid activation function. Forward update ; Using the reverse maintenance unit, according to the node Current state vector and message vector The difference calculation is performed to obtain the node. State maintenance quantity ; Using the state fusion unit, based on the node State maintenance quantity For the current state vector Modulation is performed, and the modulation result is compared with the node. message vector The candidate states are obtained by fusion. Forward update With state maintenance quantity The data is stitched together, and dynamic fusion weights are calculated using a lightweight perceptron. ; Based on dynamic fusion weights For forward update volume Modulated candidate states and state-maintaining quantities The current modulated state is weighted and fused to output the node. The updated state vector .

7. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 4, characterized in that, The step of recalculating the weights of the dynamic causal edges between nodes based on the updated state vectors of all nodes, updating the dynamic causal edge weight matrix, and obtaining the final node state vectors and the final dynamic causal edge weight matrix between each node includes: The following steps are performed through the causal structure learning unit: For any two nodes connected by a dynamic edge and nodes ,implement: Using the aforementioned causal contribution evaluation unit, the nodes are obtained respectively. and nodes The node updated in the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector ; Based on the node updated in the current iteration step State vector The node updated in the current iteration step State vector and the node in the previous iteration step. State vector With the node of the previous iteration step State vector Calculate the nodes respectively and nodes The corresponding first change Second change And use a multilayer perceptron to analyze the first change. and the second change Perform nonlinear feature mapping to obtain higher-order feature representations of the first and second transform quantities; The higher-order feature representations of the first and second transform quantities are concatenated, and the nodes are calculated based on the concatenation result. For nodes causal effect score ; Using the static prior injection unit, nodes in the static graph structure are obtained. and nodes Corresponding initial weights The initial weights With the causal effect score Perform linear weighted fusion to obtain the fusion score. ; Using the dynamic sparsity processing unit, the fusion score is... Normalize and apply a learnable threshold. The sparsification function will be below the threshold. The fusion score is set to zero, resulting in the node. and New weights of dynamic causal edges To update the dynamic causal edge weight matrix.

8. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 1, characterized in that, Fault prediction estimation includes at least one of the following: the confidence level of failure occurrence of the specified equipment, the equipment health score, and the potential root cause nodes and impact paths obtained by backtracking analysis based on the final dynamic causal edge weight matrix.

9. The intelligent grounding fault diagnosis method for photovoltaic power plants according to claim 1, characterized in that, The training process of the state reasoning network includes: Collect historical operating data and corresponding historical fault tag data of photovoltaic power plants; Construct a historical graph structure corresponding to the historical operational data; Using the historical graph structure and historical operation data as model inputs, and the historical fault label data as supervision targets, the state reasoning network, which includes the multi-hop perception aggregation unit, bidirectional information fusion unit, and causal structure learning unit, is trained end-to-end to minimize the loss function between the predicted results and the true labels.

10. An intelligent grounding fault diagnosis device for photovoltaic power plants, characterized in that, include: The data acquisition module is used to acquire real-time operating data of the photovoltaic power station; wherein, the real-time operating data includes at least: equipment status data characterizing the equipment status and environmental status data characterizing the environmental conditions; The graph structure confirmation module is used to determine the static graph structure based on the physical connection relationship and electrical logic relationship of the photovoltaic power station; wherein, the nodes in the static graph structure represent the equipment in the photovoltaic power station; The data conversion module is used to convert the real-time running data using a graph structure encoder to obtain the initial feature vectors corresponding to each node; wherein, the initial feature vectors include: a device state embedding vector representing the device state and an environmental state embedding vector representing the environmental conditions; The iterative inference module is used to input the initial feature vector and the static graph structure into the pre-trained state inference network, perform K-step iterative inference, and obtain the final node state vector and the final dynamic causal edge weight matrix between each node. The state inference network outputs the final node state vector and the final dynamic causal edge weight matrix between each node through K-step iterative inference. In each iteration, the following operations are performed sequentially: based on the dynamic causal edge weight matrix of the current iteration step, the state vectors of each node's neighboring nodes are weighted and summed to generate the node's message vector; the current state vector of each node is fused with its corresponding message vector using an update function to obtain the updated state vector of the node; based on the updated state vectors of all nodes, the weights of the dynamic causal edges between nodes are recalculated, and the dynamic causal edge weight matrix is ​​updated to obtain the final node state vector and the final dynamic causal edge weight matrix between each node. The fault prediction module is used to obtain fault prediction estimates based on the final node state vector and the final dynamic causal edge weight matrix between each node.

Images