Intelligent fault diagnosis method and system for photovoltaic system

Abstract

The invention provides an intelligent fault diagnosis method and system for a photovoltaic system. The method comprises the following steps: collecting multi-source heterogeneous operation parameter data of the photovoltaic system based on an intelligent sensor network and carrying out standardized preprocessing; local feature extraction is carried out on edge nodes, a frequency moment is calculated, data summary aggregation is carried out through a frequency moment optimization algorithm, and a fusion key feature vector set is constructed by using a multi-modal data fusion model; performing anomaly detection based on the normal behavior mode baseline, and generating an abnormal state alarm signal and a preliminary anomaly type indication; constructing a weighted graph model based on a system topological structure, constructing a fault association sub-graph by using an approximate spanning tree algorithm, accurately positioning a fault unit and diagnosing a fault type; and generating and executing a self-healing operation scheme according to the fault diagnosis conclusion. Through the distributed multi-modal data fusion and approximate spanning tree fault positioning technology, rapid and accurate diagnosis and autonomous recovery of the photovoltaic system fault are realized, and the system reliability and the operation efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, specifically to a method and system for intelligent fault diagnosis of photovoltaic systems, and particularly to a technical solution for real-time monitoring, accurate fault diagnosis and intelligent self-healing of large-scale distributed photovoltaic systems by utilizing distributed multimodal data fusion and graph theory-based fault location technology. Background Technology

[0002] Photovoltaic power generation, as a clean energy technology, has been widely applied and rapidly developed globally. However, photovoltaic systems, especially large-scale distributed photovoltaic power plants, have a large number of components and complex deployment environments, making them prone to various failures during long-term operation. Therefore, efficient and accurate fault diagnosis and timely handling of photovoltaic systems are crucial for ensuring stable system operation, improving power generation efficiency, and extending equipment lifespan. Technical research in this field mainly focuses on how to quickly detect faults, accurately locate fault sources, and intelligently perform fault recovery.

[0003] Currently, various technical methods exist in the field of photovoltaic system fault diagnosis. More basic methods include those based on electrical parameter thresholds, such as monitoring the output voltage and current of the photovoltaic array to determine if a fault has occurred. Another common method relies on manual inspection and experience-based judgment, with maintenance personnel periodically checking equipment status on-site or analyzing fault causes based on historical experience. Furthermore, some research has begun to incorporate machine learning algorithms, such as Support Vector Machines (SVM) and neural networks, to train fault diagnosis models by analyzing historical operating data, aiming to achieve more intelligent fault identification.

[0004] A relevant prior art technology involves using sensors to collect operational data from key nodes of a photovoltaic system and then performing fault diagnosis through a centralized data analysis platform. This type of technology typically monitors parameters such as the output characteristics of the photovoltaic array and the operating status of the inverter. When abnormal data is detected, the system attempts to match it with preset fault modes to determine the possible fault type. For example, it analyzes the distortion of the IV characteristic curve to determine if there are hot spots or shading on the modules, or it monitors the inverter's conversion efficiency and alarm signals to determine inverter faults. These systems improve the automation level of fault diagnosis to a certain extent.

[0005] However, the aforementioned most relevant existing technologies still have some shortcomings. First, when processing multi-source heterogeneous data (such as electrical parameters, environmental parameters, and equipment status information) in large-scale distributed photovoltaic systems, existing methods often struggle to efficiently integrate and comprehensively analyze this information. This results in room for improvement in the accuracy of fault diagnosis and the precision of fault location, especially when faced with complex system topologies and fault coupling situations, making it difficult to quickly and accurately trace the source. Second, many existing fault diagnosis models, especially in distributed deployment scenarios, have high computational overhead for data processing and model optimization. This makes it difficult to meet the high real-time requirements of photovoltaic systems for fault diagnosis and to efficiently adapt to the resource limitations of edge computing devices, leading to response delays or limited diagnostic capabilities. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent fault diagnosis method and system for photovoltaic systems, aiming to solve the technical problems of low fusion efficiency, insufficient fault location accuracy, and high computational overhead of diagnostic models when processing multi-source heterogeneous data of large-scale distributed photovoltaic systems.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0008] A method for intelligent fault diagnosis of photovoltaic systems, comprising:

[0009] Based on the intelligent sensor network deployed at each key node of the photovoltaic system, multi-source heterogeneous operating parameter data of the photovoltaic system are collected, and the multi-source heterogeneous operating parameter data is standardized and preprocessed to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

[0010] Local features are extracted and frequency moments are calculated from the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes. Data summary aggregation is performed through the frequency moment optimization algorithm. The data summary is processed using a multimodal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system.

[0011] Anomaly detection is performed on the fused key feature vector set based on a pre-trained baseline of normal behavior patterns. When an anomaly is detected, an anomaly state alarm signal, related anomaly feature data fragments, and preliminary anomaly type indication are generated.

[0012] Based on the abnormal state alarm signals, abnormal feature data fragments, and preliminary abnormal type indications, a weighted graph model is constructed based on the physical topology and electrical connection information of the photovoltaic system. An approximate spanning tree algorithm is used to construct a fault association subgraph. By analyzing the fault association subgraph, accurate fault unit location information, detailed fault type diagnosis conclusions, and an assessment report on the fault impact range and severity are output.

[0013] Based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact, a self-healing operation plan is generated. The control module then issues operation instructions to the relevant actuators and executes the self-healing operation.

[0014] Preferably, the multi-source heterogeneous operating parameter data undergoes standardized preprocessing, including:

[0015] Data cleaning is performed on the multi-source heterogeneous operating parameter data to identify and process garbled data caused by communication errors, extreme outliers exceeding the nominal value by ±50%, and sudden noise data that deviates from the mean by more than 3 times the standard deviation for three consecutive sampling periods, resulting in a pre-cleaned raw dataset.

[0016] The original dataset that has been preliminarily cleaned is checked for missing data. Linear interpolation is used for environmental parameters whose changes are below a preset threshold. Interpolation based on time-series correlation is used for periodically changing power generation parameters. Local mean interpolation is used for short-term missing data to obtain a complete original dataset.

[0017] The original dataset is timestamped and resampled. A 5-minute sampling frequency is selected as the baseline sampling frequency. High-frequency data is downsampled using a moving average method, and low-frequency data is upsampled using piecewise linear interpolation to obtain a time-aligned multi-source dataset.

[0018] The time-aligned multi-source dataset is subjected to unit normalization and format unification operations. Data with different physical dimensions are scaled to the [0,1] interval using the min-max normalization method and converted into a unified data structure to obtain the standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

[0019] Preferably, the local feature extraction and frequency moment calculation of the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes include:

[0020] Within a preset 10-minute time window, the statistical characteristics of the standardized multi-source heterogeneous photovoltaic system operating parameter data stream are calculated. The first and second moments are updated using a recursive calculation method to obtain a distributed dataset containing local features and their frequency moments. The statistical characteristics include mean, standard deviation, peak factor, skewness, and kurtosis.

[0021] For the distributed dataset containing local features and their frequency moments, a probabilistic data structure based on Count-MinSketch is applied for data compression. A hierarchical aggregation strategy is used to transfer local statistics between nodes to achieve distributed estimation of frequency moments, thereby obtaining an optimized frequency moment summary of the global data distribution.

[0022] Preferably, the data summary is processed using a multimodal data fusion model, including:

[0023] Based on the distributed dataset containing local features and the optimized frequency moment summary of the global data distribution, a multimodal data fusion model is constructed, which includes a multilayer perceptron with an encoding layer and a decoding layer, a bidirectional long short-term memory network for processing time-series data, and a three-layer convolutional neural network for feature extraction. The multimodal data fusion model is trained using a cross-validation method to obtain a trained multimodal data fusion model.

[0024] Using the trained multimodal data fusion model, the standardized multi-source heterogeneous photovoltaic system operation parameter data stream that is input in real time is processed to extract low-dimensional feature representations that characterize the current overall operation status of the photovoltaic system, thereby obtaining the fusion key feature vector set containing comprehensive operation status information of the photovoltaic system.

[0025] Preferably, anomaly detection is performed on the fused key feature vector set based on a pre-trained baseline of normal behavior patterns, including:

[0026] Based on the historical fusion key feature vector set during the normal operation of the photovoltaic system, an autoencoder model containing three encoding layers and three decoding layers is constructed. By minimizing the reconstruction error, a compact representation of the normal data is learned, and a baseline model representing the normal operation state of the photovoltaic system is obtained.

[0027] Each feature vector in the fusion key feature vector set is compared with the baseline model to calculate the magnitude of the reconstruction error and obtain the quantized real-time state deviation value.

[0028] Based on the statistical distribution of historical operating data, a 95% confidence interval is set as the early warning threshold and a 99% confidence interval is set as the alarm threshold. When the deviation value exceeds the early warning threshold for three consecutive sampling periods or exceeds the alarm threshold once, the abnormal state alarm signal and related abnormal feature data fragments are generated.

[0029] Based on the abnormal state alarm signal and related abnormal feature data fragments, the pattern of the abnormal features is matched with a predefined macroscopic abnormal category to obtain the preliminary abnormal type indication.

[0030] Preferably, a weighted graph model is constructed based on the physical topology information and electrical connection information of the photovoltaic system, including:

[0031] The complete physical topology information and detailed electrical connection information of the photovoltaic system are loaded from the preset database. The information is then cleaned and converted to obtain the structured topology data of the current photovoltaic system.

[0032] Based on the structured topology data, a node feature matrix is ​​established, and each key device in the photovoltaic system is abstracted as a node of the graph. The operating parameters of the devices are mapped as feature vectors of the nodes to obtain the initialized graph structure.

[0033] Based on the initialized graph structure, the physical connection relationships between devices are transformed into an adjacency matrix, and the electrical influence relationships are transformed into association strengths, thus obtaining the basic connection structure of the graph;

[0034] Based on the abnormal feature data fragments and preliminary abnormal type indications, the correlation coefficients and fault propagation probabilities between nodes in the graph structure are calculated, and the correlation coefficients and fault propagation probabilities between nodes in the graph structure are used as the weight attributes of the edges to obtain the edge weight matrix.

[0035] By combining the node feature matrix, basic connection structure, and edge weight matrix, a weighted graph model representing the complete topology and fault correlation of the system is constructed.

[0036] Preferably, by analyzing the fault correlation subgraph, accurate fault unit location information, detailed fault type diagnosis conclusions, and an assessment report on the fault impact range and severity are output, including:

[0037] A depth-first traversal is performed on the nodes in the fault association subgraph to extract the historical data of each node in the 10 sampling periods before the fault occurred and the real-time data after the fault occurred. Parameter trend comparison analysis is performed to obtain the time-series evolution characteristics of parameter anomalies.

[0038] Based on the temporal evolution characteristics of the aforementioned parameter anomalies, a causal relationship graph between nodes is constructed, and the correlation strength between nodes is calculated using the maximum information coefficient to obtain the critical path of fault propagation.

[0039] Based on the critical path of fault propagation, and combined with the electrical logic relationship of the equipment, a backtracking analysis is performed to locate the fault source equipment and obtain precise fault unit location information.

[0040] The precise fault unit location information is associated with relevant abnormal feature data fragments, a pre-set detailed fault knowledge base is queried, and a pattern matching method combining rule-based expert system and support vector machine is used to obtain detailed fault type diagnosis conclusions.

[0041] Based on the fault type diagnosis conclusions and fault unit location information, combined with the system topology, the Monte Carlo simulation method is used to evaluate the propagation range and impact of the fault at different time scales, and an assessment report on the impact range and severity of the fault is obtained.

[0042] Preferably, the self-healing operation plan includes:

[0043] Historical processing cases related to the fault type diagnosis conclusion are extracted from the preset self-healing strategy library, and similarity calculation and strategy matching are performed using the case reasoning method to obtain the initial candidate self-healing operation scheme.

[0044] Using a deep reinforcement learning model, a policy network is trained based on historical operating data to generate new candidate processing schemes for the current fault scenario. These schemes are then merged with the initial candidate self-healing operation schemes to obtain a complete set of candidate self-healing operation schemes.

[0045] A multi-objective evaluation model is constructed, which uses factors such as expected fault recovery time, system performance recovery degree, operational complexity, resource consumption and potential risks as evaluation indicators to comprehensively score the candidate self-healing operation scheme set and obtain the scheme evaluation results.

[0046] Based on the evaluation results of the proposed schemes, the weight of each evaluation index is determined using the analytic hierarchy process (AHP), the comprehensive score of each scheme is calculated, and the scheme with the highest score is selected as the optimal self-healing operation scheme.

[0047] The optimal self-healing operation scheme is decomposed into a sequence of control instructions at the device level, and an instruction execution flowchart is established based on the execution sequence dependency to obtain the operation instructions.

[0048] Preferably, the method further includes: monitoring system status parameters after the self-healing operation, including:

[0049] A multi-dimensional monitoring index system including power generation efficiency, equipment temperature, electrical parameters and communication status is established, and real-time data is collected on key nodes in the self-healing operation process to obtain the status monitoring data stream of the execution process;

[0050] The status monitoring data stream is compared with the historical baseline data before the failure, and the recovery degree of each monitoring indicator is calculated. When the key indicators recover to more than 90% of the baseline value, the self-healing operation is confirmed to be successful, and real-time data on the execution status of the self-healing operation and the system performance recovery status are obtained.

[0051] A hierarchical log recording model is constructed, and information such as alarm information, diagnosis process, self-healing measures, operation response and performance recovery are organized according to time sequence and logical relationship to generate a structured event log, thus obtaining the log record containing the entire fault diagnosis and self-healing process.

[0052] Extract key information on fault characteristics, diagnostic rules, and self-healing strategies from the log records, perform knowledge extraction and rule summarization, and establish a knowledge graph update strategy.

[0053] When the self-healing operation is successful and the performance recovery exceeds expectations, the extracted knowledge and rules are updated to the fault knowledge base and self-healing strategy base to optimize the system's diagnostic and processing capabilities.

[0054] A photovoltaic system intelligent fault diagnosis system, comprising:

[0055] The acquisition module is used to acquire multi-source heterogeneous operating parameter data of the photovoltaic system based on the intelligent sensor network deployed at each key node of the photovoltaic system, and to perform standardized preprocessing on the multi-source heterogeneous operating parameter data to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

[0056] The calculation module is used to extract local features and calculate frequency moments of the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes, perform data summary aggregation through the frequency moment optimization algorithm, and process the data summary using a multi-modal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system.

[0057] An anomaly detection module is used to perform anomaly detection on the fused key feature vector set based on a pre-trained normal behavior pattern baseline. When an anomaly is detected, it generates an anomaly state alarm signal, related anomaly feature data fragments, and a preliminary anomaly type indication.

[0058] The fault diagnosis module is used to construct a weighted graph model based on the physical topology and electrical connection information of the photovoltaic system, according to the abnormal state alarm signal, abnormal feature data fragments and preliminary abnormal type indication, and to construct a fault association subgraph using an approximate spanning tree algorithm. By analyzing the fault association subgraph, it outputs accurate fault unit location information, detailed fault type diagnosis conclusions and an assessment report on the fault impact range and severity.

[0059] The self-healing operation module is used to generate a self-healing operation plan based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact. It then issues operation instructions to relevant actuators through the control module and executes the self-healing operation.

[0060] The beneficial effects of this invention are:

[0061] By using a distributed multimodal data fusion method based on frequency moment optimization, local feature extraction and frequency moment calculation are performed at edge nodes, which reduces the amount of data transmission and improves data processing efficiency. At the same time, it can effectively fuse multi-source heterogeneous data to construct a key feature vector that comprehensively characterizes the operating status of photovoltaic systems.

[0062] A fault location method based on approximate spanning tree is adopted. By combining system topology information and abnormal features, a fault association subgraph is quickly constructed, which realizes the accurate location of fault root causes and the identification of detailed fault types, thereby improving the accuracy and efficiency of fault diagnosis.

[0063] A complete closed-loop system was constructed, encompassing data acquisition, anomaly detection, fault diagnosis, and self-healing operation. This system enables intelligent and autonomous handling of photovoltaic system faults, improving system reliability and operational efficiency while reducing the need for manual intervention. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of the intelligent fault diagnosis method for photovoltaic systems of the present invention;

[0066] Figure 2 This is a flowchart of the real-time acquisition and standardized preprocessing of multi-source heterogeneous data in this invention;

[0067] Figure 3 This is a flowchart illustrating the distributed multimodal data fusion and key feature vector construction based on frequency moment optimization according to the present invention.

[0068] Figure 4 This is a structural block diagram of the intelligent fault diagnosis system for photovoltaic systems of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0070] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0071] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0072] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] like Figure 1 As shown, the intelligent fault diagnosis method for photovoltaic systems provided by this invention includes the following steps:

[0074] S1. Based on the intelligent sensor network deployed at each key node of the photovoltaic system, collect multi-source heterogeneous operating parameter data of the photovoltaic system, perform standardized preprocessing on the multi-source heterogeneous operating parameter data, and obtain standardized multi-source heterogeneous photovoltaic system operating parameter data stream;

[0075] S2. At the edge nodes, local features are extracted from the data stream of the operating parameters of the multi-source heterogeneous photovoltaic system and frequency moments are calculated. Data summary aggregation is performed through the frequency moment optimization algorithm. The data summary is processed using a multimodal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system.

[0076] S3. Based on the pre-trained normal behavior pattern baseline, perform anomaly detection on the fused key feature vector set. When an anomaly is detected, generate an anomaly state alarm signal, related anomaly feature data fragments, and preliminary anomaly type indication.

[0077] S4. Based on the abnormal state alarm signal, abnormal feature data fragments, and preliminary abnormal type indication, construct a weighted graph model based on the physical topology information and electrical connection relationship information of the photovoltaic system, construct a fault association subgraph using the approximate spanning tree algorithm, and output accurate fault unit location information, detailed fault type diagnosis conclusions, and an assessment report on the fault impact range and severity by analyzing the fault association subgraph.

[0078] S5. Based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact, generate a self-healing operation plan, and issue operation instructions to relevant actuators through the control module to execute the self-healing operation.

[0079] The detailed explanation of the above steps is as follows:

[0080] Step S1: Multi-source heterogeneous data acquisition and standardization preprocessing

[0081] This step establishes the basic architecture for data acquisition in the photovoltaic system. The system deploys an intelligent sensor network at various key nodes of the photovoltaic power plant, including each string level of the photovoltaic array, combiner boxes, inverter input / output terminals, power distribution equipment connection points, and environmental monitoring points. These sensors encompass various types, including voltage sensors, current sensors, temperature sensors, irradiance sensors, wind speed sensors, and humidity sensors, enabling comprehensive collection of electrical parameters, environmental parameters, and equipment status information of the photovoltaic system.

[0082] The data standardization preprocessing process comprises four key steps. First, the data cleaning stage identifies and processes garbled data caused by communication errors, extreme outliers exceeding ±50% of the nominal value, and sudden noise data deviating from the mean by more than three standard deviations for three consecutive sampling periods. Second, missing data imputation employs appropriate imputation strategies for different types of parameters: linear imputation is used for slowly changing environmental parameters, time-series correlation imputation based on the ARIMA model is used for power generation parameters with periodic variations, and local mean imputation is used for short-term missing data. Third, timestamp alignment and resampling are performed, using a uniform 5-minute sampling frequency. High-frequency data is downsampled using a moving average method, while low-frequency data is upsampled using piecewise linear interpolation. Finally, data normalization and format unification are implemented, using a min-max normalization method to scale data with different physical dimensions to the [0,1] interval, forming a unified data structure.

[0083] Step S2: Distributed multimodal data fusion and feature vector construction

[0084] This step enables efficient data processing within a distributed edge computing architecture. At each edge node, the system extracts local features from the standardized data stream within a 10-minute time window, calculating statistical features including mean, standard deviation, kurtosis, skewness, and kurtosis, and uses a recursive calculation method to update the first and second moments in real time. To reduce data transmission burden and improve processing efficiency, the system applies a probabilistic data structure based on Count-Min Sketch for data compression. This structure uses five hash functions and a 1024×5 counting matrix, and transmits local statistics between nodes through a hierarchical aggregation strategy to achieve distributed estimation of frequency moments.

[0085] The multimodal data fusion model employs a deep learning architecture, comprising three main components working collaboratively. The multilayer perceptron component handles static features, consisting of an encoding layer (256-128-64 neurons) and a decoding layer (64-128-256 neurons), with batch normalization and a dropout layer of 0.3 added after each layer. The bidirectional long short-term memory network component specifically processes temporal data, employing a two-layer structure (128 and 64 LSTM units) capable of capturing sequence patterns over 12 historical time steps. The convolutional neural network component is used for spatial feature extraction, employing a three-layer convolutional structure (32-64-128 convolutional kernels), progressively extracting abstract features through 3×3 convolutional kernels and 2×2 max-pooling layers. The outputs of these three components are concatenated through a fusion layer, ultimately generating a 32-dimensional fused key feature vector, comprehensively representing the operating status of the photovoltaic system.

[0086] Step S3: Intelligent Anomaly Detection Based on Behavioral Baselines

[0087] This step establishes a robust anomaly detection mechanism. The system first constructs a baseline of normal behavior patterns and then uses an autoencoder model to learn a compact representation of normal operating data. This autoencoder contains three encoding layers (24-16-8 neurons) and three decoding layers (16-24-32 neurons), learning the inherent patterns of normal data by minimizing the reconstruction error. The model is trained using approximately 500,000 normal operating data points, employing the Adam optimizer with a learning rate of 0.001, and incorporating L2 regularization to prevent overfitting.

[0088] The anomaly detection process employs a statistical threshold determination method. The system inputs the real-time generated fused key feature vector into the baseline model and calculates the reconstruction error as the state deviation value. Based on the statistical distribution of reconstruction errors from historical normal data, a 95% confidence interval (threshold 0.022) is set as the warning threshold, and a 99% confidence interval (threshold 0.035) is set as the alarm threshold. When the deviation value exceeds the warning threshold for three consecutive sampling periods or exceeds the alarm threshold once, the system generates an abnormal state alarm signal and records the feature data of the preceding and following 15 sampling periods as anomaly feature data fragments. By performing cosine similarity matching with predefined macroscopic anomaly categories, the system can provide preliminary anomaly type indications, providing directional guidance for subsequent accurate diagnosis.

[0089] Step S4: Precise Fault Location and Diagnosis Based on Graph Theory

[0090] This step establishes an intelligent fault diagnosis mechanism based on system topology. The system first loads the complete physical topology and electrical connection information of the photovoltaic system from the configuration database, including the series and parallel connections of components, combiner box connections, inverter configurations, etc., and constructs a weighted graph model. In the graph model, each device is abstracted as a node, and node features include 20-30 dimensions of information such as device type, geographical location, operating parameters, and historical fault frequency. Edge weights are calculated using three factors: the correlation strength of electrical influence relationships, the correlation coefficient calculated based on anomaly features, and the fault propagation probability estimated based on historical patterns, with weight coefficients of 0.3, 0.4, and 0.3, respectively.

[0091] Fault localization employs an approximate spanning tree algorithm for optimization. Starting from anomaly-related seed nodes, the system iteratively constructs a fault-related subgraph by selecting edges with higher weights. A depth-first traversal is performed on the nodes in the subgraph to extract time-series data before and after the fault occurrence for trend analysis, constructing a causal relationship graph between nodes. The maximum information coefficient (MIC) is used to calculate the correlation strength between nodes; causal edges are established when the MIC value exceeds 0.7, with the edge direction determined based on temporal precedence. By finding the node with the minimum in-degree in the causal graph and combining this with backtracking analysis of electrical logic relationships, the system accurately locates the fault-originating device.

[0092] Fault type diagnosis combines expert systems and machine learning methods. The system queries a fault knowledge base containing approximately 200 rules, uses a rule-based expert system for initial matching, and then uses a trained support vector machine model (RBF kernel, C=10, gamma=0.1) for fine-grained classification. Fault impact assessment uses Monte Carlo simulation, running 1000 simulations, and uses a Markov chain model to predict the propagation range and impact of the fault in the short term (1 hour), medium term (24 hours), and long term (7 days), generating a comprehensive assessment report including power generation loss, equipment damage risk, and system stability impact.

[0093] Step S5: Generation and Execution of Intelligent Self-Healing Strategy

[0094] This step implements a closed-loop intelligent self-healing mechanism. The self-healing solution generation employs a dual strategy: a combination of case reasoning and reinforcement learning. The system extracts historical processing cases related to the current fault type from the self-healing policy library, calculates similarity using weighted Euclidean distance, and selects the top 5 most similar cases as initial candidate solutions. Simultaneously, a deep reinforcement learning model is used to generate new candidate solutions. This model is trained using the PPO algorithm and includes a policy network (4 fully connected layers with 256-128-64-32 hidden neurons) and a value network (3 fully connected layers with 128-64 hidden neurons), trained for 10,000 rounds in a simulation environment built based on historical data.

[0095] The solution evaluation employs a multi-objective decision analysis method. The system constructs an evaluation index system encompassing five dimensions: fault recovery time, performance recovery degree, operational complexity, resource consumption, and potential risks. The weight vector (0.15, 0.35, 0.20, 0.10, 0.20) is determined using the analytic hierarchy process. Each candidate solution is quantitatively evaluated according to preset scoring rules, and a weighted comprehensive score is calculated. The solution with the highest score is selected as the optimal self-healing solution.

[0096] Operation execution comprises three stages: instruction decomposition, process planning, and real-time monitoring. The system decomposes the advanced self-healing scheme into basic operational units, establishes a directed acyclic graph to represent operational dependencies, and identifies critical paths and parallel execution branches. Operation instructions are converted into control commands recognizable by the equipment (such as Modbus commands and IEC 61850 messages) and sent to the actuators via industrial Ethernet or wireless sensor networks. During execution, a multi-dimensional monitoring indicator system is established, collecting data such as power generation efficiency, equipment temperature, electrical parameters, and communication status in real time, and comparing this data with historical baseline data to calculate the degree of recovery. Self-healing is confirmed as successful when key indicators recover to more than 90% of the baseline value. The system generates hierarchical log records, extracting key information such as fault characteristics, diagnostic rules, and self-healing strategies. When the self-healing effect exceeds expectations, the knowledge base and strategy base are updated to achieve continuous system optimization.

[0097] In one embodiment, step S1: Real-time acquisition and standardized preprocessing of multi-source heterogeneous data from the distributed photovoltaic system. For example... Figure 2 As shown, this step specifically includes:

[0098] Step S11: Deployment and Synchronous Data Acquisition of Multi-Source Sensors. Based on the intelligent sensor network deployed at each key node of the photovoltaic system, multi-source heterogeneous operating parameter data of the photovoltaic system are collected.

[0099] Specifically, smart sensors (including voltage sensors, current sensors, temperature sensors, irradiance sensors, wind speed sensors, humidity sensors, and equipment status monitoring modules) are pre-deployed in the physical layout of the photovoltaic system and key equipment (such as each string or combiner box level of the photovoltaic array, the input and output terminals of the inverter, the connection points of key power distribution equipment, and representative environmental monitoring points). During data acquisition, a unified clock synchronization mechanism (such as NTP or PTP protocol) is used to ensure high-precision consistency of the timestamps of each sensor data, thereby obtaining raw multi-source heterogeneous data packets with accurate timestamps.

[0100] Step S12: Raw data cleaning and outlier handling. The multi-source heterogeneous operating parameter data is cleaned to identify and process garbled data caused by communication errors, extreme outliers exceeding ±50% of the nominal value, and sudden noise data that deviates from the mean by more than 3 times the standard deviation for three consecutive sampling periods, resulting in a pre-cleaned raw dataset.

[0101] In practice, the system first sets reasonable physical ranges for various parameters. For example, the output voltage of photovoltaic modules is usually between 0-50V. When a voltage value exceeding 75V or falling below -25V is detected (i.e., exceeding the nominal value by ±50%), it is marked as an extreme outlier and removed. For sudden noise data, the system calculates the mean and standard deviation of each parameter within a certain time window (e.g., 30 minutes). When a data point deviates from the mean by more than three times the standard deviation for three consecutive sampling periods, it is identified as noise data and smoothed.

[0102] Step S13: Missing Data Imputation. The pre-cleaned raw dataset is checked for missing data. Linear imputation is used for environmental parameters with changes below a preset threshold, time-series correlation-based imputation is used for periodically changing power generation parameters, and local mean imputation is used for short-term missing data, resulting in a complete raw dataset.

[0103] Specifically, the system first detects missing points in the data stream. For parameters with slow changes, such as ambient temperature (e.g., changes of no more than 0.5℃ within 5 minutes), a linear interpolation method is used to fill in the missing values. For parameters with obvious periodic changes, such as solar irradiance and power generation, the system uses a time-series interpolation method based on the ARIMA (Autoregressive Integral Moving Average) model. The model parameters are set to p=2 (number of autoregressive terms), d=1 (difference times), and q=1 (number of moving average terms). The system predicts reasonable values ​​for missing points by analyzing the historical change patterns of the parameters. For short-term (no more than 3 sampling periods) random missing data, the local mean of the five valid data points before and after the missing data is used for interpolation.

[0104] Step S14: Data Timestamp Alignment and Resampling. The complete original dataset is timestamped and resampled. A 5-minute sampling frequency is selected as the baseline. High-frequency data is downsampled using a moving average method, and low-frequency data is upsampled using piecewise linear interpolation, resulting in a time-aligned multi-source dataset.

[0105] In practice, the system first standardizes all data to a 5-minute time interval. For data with a sampling frequency higher than 5 minutes (such as current data sampled every 1 minute), a 5-minute window sliding average is used for downsampling. For data with a sampling frequency lower than 5 minutes (such as environmental data sampled every 15 minutes), a piecewise linear interpolation method is used for upsampling.

[0106] Step S15: Data Unit Normalization and Format Unification. The time-aligned multi-source dataset is normalized and format unified. Data with different physical dimensions are scaled to the [0,1] interval using a min-max normalization method, converting them into a unified data structure to obtain the standardized multi-source heterogeneous photovoltaic system operating parameter data stream. Specifically, for each physical quantity (such as voltage, current, temperature, etc.), the system first determines its theoretical or historical observation minimum value (min). i and maximum value max i Then, for each original value x, apply the min-max normalization formula: x' = (x - min i ) / (max i -min i This maps the temperature of a photovoltaic module to the interval [0,1]. For example, for the temperature of a photovoltaic module, the minimum value is taken. i = -10℃ (lowest winter temperature) and max i =85℃ (extreme high temperature condition), then the normalized value of the original temperature of 35℃ is (35-(-10)) / (85-(-10)) = 0.47. The normalized data is organized into a unified structured array or DataFrame format, with each row representing a time point and each column representing a sensor parameter, forming a standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

[0107] Step S2: Distributed multimodal data fusion and key feature vector construction based on frequency moment optimization.

[0108] like Figure 3 As shown, this step specifically includes:

[0109] Step S21: Local Feature Extraction and Preliminary Calculation of Frequency Moments at Edge Nodes. Local feature extraction and frequency moment calculation are performed on the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes.

[0110] Specifically, within a preset 10-minute time window, the statistical characteristics of the standardized multi-source heterogeneous photovoltaic system operating parameter data stream are calculated. A recursive calculation method is used to update the first and second moments, resulting in a distributed dataset containing local features and their frequency moments. These statistical features include mean, standard deviation, peak factor, skewness, and kurtosis. At each edge node (such as an edge computing device installed near a combiner box or inverter), the system extracts local features from the data of the equipment it monitors (such as the voltage and current data of a photovoltaic subarray).

[0111] These statistical features and their frequency moment information constitute a distributed dataset containing local features and their frequency moment information.

[0112] Step S22: Data summary information aggregation based on frequency moment optimization algorithm. For the distributed dataset containing local features and their frequency moment information, a probabilistic data structure based on Count-Min Sketch is applied for data compression. A hierarchical aggregation strategy is used to transfer local statistics between nodes to achieve distributed estimation of frequency moments, thereby obtaining an optimized frequency moment summary of the global data distribution.

[0113] In its implementation, the system employs a Count-Min Sketch data structure, which uses d hash functions and a w×d counting matrix, where d = 5 and w = 1024. The hash functions utilize a common hash family. Each edge node maps its local statistical features into the Count-Min Sketch structure, and then these compressed statistics are transmitted and merged between nodes using a hierarchical aggregation strategy (such as tree aggregation or gossip protocol). The merging operation involves adding the count values ​​at corresponding positions. This method significantly reduces the amount of data that needs to be transmitted (from O(n) of the original data to O(w×d)) while maintaining a good approximation of the frequency distribution. Finally, the central node or the collaborating edge node network can recover an optimized frequency moment summary of the global data distribution from the merged Count-Min Sketch, including statistics such as global mean, variance, skewness, and kurtosis.

[0114] Step S23: Construction and Training of Multimodal Data Fusion Model. Based on the distributed dataset containing local features and the optimized frequency moment summary of the global data distribution, a multimodal data fusion model is constructed, comprising a multilayer perceptron with encoding and decoding layers, a bidirectional long short-term memory network for processing time-series data, and a three-layer convolutional neural network for feature extraction. The multimodal data fusion model is trained using cross-validation to obtain the trained multimodal data fusion model.

[0115] Specifically, the multimodal data fusion model constructed by the system includes three main components:

[0116] Multilayer Perceptron (MLP) Components:

[0117] Input layer: Dimension equals the number of original features (e.g., 100 sensor parameters)

[0118] Coding layer: Three-layer structure with 256, 128, and 64 neurons respectively, using ReLU activation function.

[0119] Decoding layer: Three-layer structure with 64, 128, and 256 neurons respectively, using ReLU activation function.

[0120] Batch normalization: Add a batch normalization layer after each hidden layer, with parameters momentum = 0.9 and epsilon = 1e-5.

[0121] Drop layer: A drop layer is added after each hidden layer, with a drop rate of 0.3.

[0122] Bidirectional Long Short-Term Memory (BiLSTM) network components:

[0123] Input: Sequence data from the past 12 time steps

[0124] First layer: 128 LSTM units, bidirectional processing, returning a sequence.

[0125] The second layer consists of 64 LSTM units, providing bidirectional processing and returning only the final output.

[0126] A drop-off layer is added after each layer, with a drop-off rate of 0.2%.

[0127] Convolutional Neural Network (CNN) Components:

[0128] Input: Reshape the data into an image-like format (batch_size, height, width, channels)

[0129] First convolutional layer: 32 3×3 convolutional kernels, stride 1, padding 'same', activation function is ReLU.

[0130] First pooling layer: 2×2 max pooling, with a step size of 2.

[0131] The second convolutional layer consists of 64 3×3 convolutional kernels with a stride of 1, padding of 'same', and the activation function is ReLU.

[0132] Second pooling layer: 2×2 max pooling, with a step size of 2.

[0133] The third convolutional layer consists of 128 3×3 convolutional kernels with a stride of 1, padding of 'same', and the activation function is ReLU.

[0134] Global average pooling layer: reduces the feature map to a one-dimensional vector.

[0135] Fusion layer:

[0136] Concatenate the outputs of MLP, BiLSTM, and CNN into a single vector.

[0137] Fully connected layer: 128 neurons, activation function is ReLU

[0138] Drop-off layer: Drop-off rate 0.3

[0139] Output layer: 32 neurons, activation function is a linear function.

[0140] Training parameters:

[0141] Optimizer: Adam, learning rate 0.001, beta_1 = 0.9, beta_2 = 0.999

[0142] Batch size: 64

[0143] Training rounds: 100, with early stop mechanism (patience = 10).

[0144] Loss function: Mean Squared Error (MSE)

[0145] Cross-validation: 50% cross-validation

[0146] The model was trained using a historical dataset containing data from normal operation and various failure scenarios, with a total sample size of approximately 1 million records. During training, the system monitored the loss value on the validation set, and terminated training early when the validation loss stopped decreasing after 10 consecutive epochs. The final model achieved a mean squared error of 0.015 on the test set, with R... 2 The value is 0.92, indicating that the model has good generalization ability.

[0147] Step S24: Real-time generation of fused key feature vectors. Using the trained multimodal data fusion model, the standardized multi-source heterogeneous photovoltaic system operating parameter data stream input in real time is processed to extract low-dimensional feature representations characterizing the current overall operating state of the photovoltaic system, thus obtaining the fused key feature vector set containing comprehensive operating state information of the photovoltaic system.

[0148] Specifically, the system processes the real-time data stream using the same preprocessing steps as the training data, and then inputs it into the trained multimodal fusion model. The model's output is a 32-dimensional feature vector, with each dimension representing an abstract feature of the photovoltaic system's state. These features are automatically extracted through the model's learning process and can characterize key state information of the system, such as power generation efficiency, equipment health status, and environmental adaptability. The system generates such a fused key feature vector every 5 minutes and adds it to the feature vector set for subsequent anomaly detection and fault diagnosis.

[0149] Detailed examples of steps S21-S24

[0150] Step S21: Local feature extraction and preliminary calculation of frequency moments for edge nodes

[0151] Taking a 10MW distributed photovoltaic power station as an example, the station contains 200 strings, each equipped with an edge computing node. One edge node monitors a string containing 20 photovoltaic modules. Within a 10-minute time window at 10:00 AM, this node collects 2000 data points (4 parameters per module, sampled once per minute). The system calculates the statistical characteristics of this data: the average module voltage is 35.2V, the standard deviation is 1.8V, the peak value is 1.05, the skewness is -0.15, and the kurtosis is 2.8. Through recursive calculation, the system updates the first moment (mean) and second moment (variance) in real time. It finds that the average voltage of a certain module has decreased by 8% compared to the previous time window, while the second moment of temperature has increased by 15%. These changes are recorded as local anomalies.

[0152] Edge nodes also calculate frequency moment information, statistically analyzing the frequency of occurrence across different parameter value ranges. For example, dividing the voltage range into 10 intervals reveals that the frequency in the 30-32V interval increases from the normal 5% to 15%, a change in frequency distribution that may indicate a potential problem. The system organizes these local features and frequency information into a distributed dataset, with each edge node maintaining a feature summary for its assigned region, preparing for subsequent global data fusion.

[0153] Step S22: Aggregation of data summary information based on frequency moment optimization algorithm

[0154] Continuing the example above, each of the 200 edge nodes generates a dataset containing local features. Directly transmitting the raw data to the central node would require approximately 400MB of data. The system uses a Count-Min Sketch data structure for compression, mapping each node's statistical features to a 5×1024 counting matrix, reducing the data size to only 20KB. For example, the average voltage of a node, 35.2V, is mapped to a specific location in the matrix using a hash function, with the count value at that location receiving a corresponding weight.

[0155] The hierarchical aggregation strategy employs a tree structure. First, four adjacent edge nodes form a cluster. Nodes within a cluster exchange their respective Count-Min Sketch matrices and merge them (adding the counts at corresponding positions). Then, representative nodes from the 50 clusters are further aggregated to form a global optimized frequency moment summary. In this way, the system discovered that 12% of the modules in the entire power plant had low voltage, and the temperature distribution exhibited significant spatial clustering, with module temperatures in certain areas generally 3-5 degrees Celsius higher. This global pattern information provided crucial clues for fault location.

[0156] Step S23: Construction and Training of Multimodal Data Fusion Model

[0157] The power plant's multimodal fusion model was trained using historical data from the past two years. The dataset contains 1 million records, covering various scenarios such as normal operation, equipment failure, and severe weather. The multilayer perceptron components process static features, such as instantaneous values ​​of voltage, current, and temperature for each component at a given moment. For example, the input layer receives a 100-dimensional feature vector (20 components × 5 parameters), which is compressed to 64 dimensions by the encoding layer and then recovered by the decoding layer, learning the intrinsic representation of the data.

[0158] The bidirectional long short-term memory (LSTM) network component is specifically designed to handle temporal patterns, taking as input sequential data from the past 12 time steps (i.e., 1 hour). For example, a power output sequence [980W, 985W, 990W, ..., 920W] shows an evolution from normal to abnormal, and BiLSTM can capture this temporal dependence. The convolutional neural network component reshapes sensor data into an image-like format, such as converting 10×10 sensor array data into a two-dimensional matrix. Through convolution operations, it extracts spatial correlation features and can identify anomalous patterns in local regions.

[0159] During training, the model learns to fuse information from different modalities into a unified feature representation. For example, when a region simultaneously experiences multimodal anomalies such as increased temperature, decreased voltage, and reduced power, the fused 32-dimensional feature vector can comprehensively reflect this complex anomaly state, providing a more robust feature foundation for subsequent anomaly detection.

[0160] Step S24: Real-time generation of key feature vectors

[0161] The system generates a fusion key feature vector every 5 minutes, representing the real-time operating status of the entire photovoltaic system. For example, on a sunny morning, the feature vector generated during normal operation might be [0.85, 0.92, 0.78, 0.66, ..., 0.81], where these 32 values ​​represent different abstract features of the system. The first dimension, 0.85, might represent overall power generation efficiency; the second dimension, 0.92, might represent equipment health; and the third dimension, 0.78, might reflect environmental adaptability, etc.

[0162] When a string experiences a blocking fault, the corresponding feature vector becomes [0.78, 0.89, 0.72, 0.55, ..., 0.73], with significant changes in values ​​across multiple dimensions. By continuously monitoring these feature vector changes, the system can promptly detect abnormal patterns. For example, if the dimension related to power generation efficiency is below 0.8 for three consecutive time points, while the historical normal level is usually above 0.85, the system will trigger the anomaly detection process. This high-dimensional feature representation provides a more comprehensive reflection of the system state than simply monitoring single parameters such as voltage and current, improving the accuracy and timeliness of anomaly detection.

[0163] Step S3: Intelligent anomaly detection and preliminary diagnosis based on fused key feature vectors.

[0164] This step specifically includes:

[0165] Step S31: Establishing a baseline for normal behavior patterns.

[0166] Anomaly detection is performed on the fused key feature vector set based on a pre-trained baseline of normal behavior patterns.

[0167] Specifically, based on the historical fusion of key feature vectors during the normal operation of the photovoltaic system, an autoencoder model containing three encoding layers and three decoding layers is constructed. By minimizing the reconstruction error, a compact representation of the normal data is learned, resulting in a baseline model characterizing the normal operation state of the photovoltaic system. The detailed structure of the autoencoder model is as follows:

[0168] Input layer: 32 neurons (consistent with the dimension of the fused feature vector)

[0169] Coding layer 1: 24 neurons, activation function is ReLU

[0170] Coding layer 2: 16 neurons, activation function is ReLU

[0171] Coding layer 3: 8 neurons, activation function is ReLU (bottleneck layer)

[0172] Decoding layer 1: 16 neurons, activation function is ReLU

[0173] Decoding layer 2: 24 neurons, activation function is ReLU

[0174] Decoding layer 3: 32 neurons, activation function is a linear function (output layer)

[0175] Training parameters:

[0176] Optimizer: Adam, learning rate 0.001

[0177] Batch size: 128

[0178] Training rounds: 200, with early stop mechanism (patience = 20).

[0179] Loss function: Mean Squared Error (MSE)

[0180] Regularization: L2 regularization, with a coefficient of 0.001.

[0181] The model was trained using only normal data (approximately 500,000 records), with the goal of minimizing the mean squared error between the input and reconstructed data. After training, the model's reconstruction error on normal data had a mean of approximately 0.01 and a standard deviation of approximately 0.005. These statistics were used to subsequently set the threshold for anomaly detection.

[0182] Step S32: Calculate the real-time feature vector and baseline deviation.

[0183] Each feature vector in the fused key feature vector set is compared with the baseline model to calculate the reconstruction error and obtain a quantized real-time state deviation value. Specifically, the system inputs the real-time generated fused key feature vectors into the trained autoencoder model to obtain the reconstructed feature vectors, and then calculates the mean squared error (MSE) between the input vector and the reconstructed vector as the deviation value.

[0184] Step S33: Abnormal state determination and alarm generation.

[0185] Based on the statistical distribution of historical operational data, a 95% confidence interval is set as the early warning threshold, and a 99% confidence interval is set as the alarm threshold. When the deviation value exceeds the early warning threshold for three consecutive sampling periods or exceeds the alarm threshold in a single instance, the abnormal state alarm signal and related abnormal feature data fragments are generated. Specifically, based on the statistical distribution of normal data reconstruction errors, the system calculates the threshold corresponding to the 95% confidence interval as 0.022 (i.e., the reconstruction error of 95% of the samples in normal data is less than this value), and the threshold corresponding to the 99% confidence interval as 0.035. When the reconstruction error of the real-time feature vector is detected to exceed 0.022 for three consecutive sampling periods (i.e., 15 minutes), or exceeds 0.035 in a single instance, the system determines the current state as abnormal and generates an abnormal state alarm signal. Simultaneously, the system records the fused key feature vectors for 15 sampling periods (i.e., 75 minutes) before and after the alarm trigger time as abnormal feature data fragments for subsequent analysis.

[0186] Step S34: Preliminary anomaly type indication.

[0187] Based on the abnormal state alarm signal and related abnormal feature data fragments, the system matches the abnormal feature patterns with predefined macroscopic abnormality categories to obtain the preliminary abnormality type indication. Specifically, the system predefines several common macroscopic abnormality categories, such as "significantly decreased output power," "abnormal communication data," and "voltage / current parameter exceeding limits." For each abnormality category, the system predefines its typical pattern in the fused feature space (such as abnormal change patterns in specific dimensions) through expert knowledge and historical case analysis. When an abnormality is detected, the system calculates the similarity between the abnormal feature data fragments and these predefined patterns, using cosine similarity as the metric.

[0188] Detailed examples of steps S31-S34

[0189] Step S31: Establishing a baseline for normal behavioral patterns

[0190] The photovoltaic power station accumulated 500,000 fused key feature vector data points during its normal operation over the past year, covering different seasons, weather conditions, and load conditions. The system uses this normal data to train an autoencoder model to learn a compact representation of the normal operating state. For example, the normal feature vector [0.85,0.92,0.78,0.66,...,0.81] on a sunny spring morning is compressed into an 8-dimensional bottleneck representation [0.45,0.67,0.23,0.89,0.34,0.78,0.56,0.91] by the encoding layer, and then reconstructed into [0.84,0.91,0.79,0.65,...,0.82] by the decoding layer.

[0191] The reconstruction error is calculated as the mean square error between the original vector and the reconstructed vector. Under normal circumstances, this error is very small, typically between 0.005 and 0.015. During model training, the system discovered the distribution characteristics of reconstruction errors under different operating modes: the mean reconstruction error is 0.008 under full load on sunny days, 0.012 under partial load on cloudy days, and 0.015 during the early morning startup phase. Through training with a large amount of normal data, the autoencoder learned to recognize the inherent patterns of various normal operating modes, establishing a reliable baseline reference for anomaly detection.

[0192] Step S32: Calculation of real-time feature vector and baseline deviation

[0193] When the system detects a real-time feature vector of [0.72, 0.85, 0.69, 0.48, ..., 0.65] at a certain moment, it inputs this vector into the trained autoencoder model. After an encoding-decoding process, a reconstructed vector of [0.83, 0.90, 0.77, 0.64, ..., 0.80] is obtained. Calculating the reconstruction error reveals significant differences between the original vector and the reconstructed vector in multiple dimensions. The difference is 0.11 in the first dimension and 0.16 in the fourth dimension, with an overall mean squared error of 0.032.

[0194] This error value far exceeds the normal range of 0.008-0.015, indicating that the current system state deviates from the normal operating mode. Analysis of the reconstruction error distribution across various dimensions revealed that the dimension related to power generation efficiency had the largest deviation, followed by the dimension related to equipment temperature. This deviation pattern suggests potential issues such as decreased power generation efficiency and equipment overheating. The system records this 0.032 deviation value as a quantified real-time state deviation value for subsequent anomaly detection.

[0195] Step S33: Abnormal State Determination and Alarm Generation

[0196] Based on the statistical distribution analysis of reconstruction errors from historical normal data, the system determined the threshold corresponding to the 95% confidence interval to be 0.022 and the threshold corresponding to the 99% confidence interval to be 0.035. When a reconstruction error of 0.032 was detected, although it did not exceed the alarm threshold of 0.035, it exceeded the warning threshold of 0.022. The system began to closely monitor subsequent data. In the next two sampling periods (after 10 minutes and 15 minutes), the reconstruction errors were 0.029 and 0.031, respectively, exceeding the warning threshold for three consecutive sampling periods.

[0197] According to the judgment rules, the system immediately generates an abnormal state alarm signal and records 15 fused key feature vectors as abnormal feature data segments within a 75-minute time period, from 30 minutes before the anomaly occurred to 45 minutes after the anomaly was confirmed. These data show the complete process of the state gradually evolving from normal [0.85, 0.92, 0.78, ...] to abnormal [0.72, 0.85, 0.69, ...], providing valuable time-series information for subsequent fault analysis. The alarm signal simultaneously triggered a red warning display on the on-site monitoring screen and an SMS notification to maintenance personnel.

[0198] Step S34: Preliminary anomaly type indication

[0199] The system predefines five common macroscopic anomaly categories and their typical feature patterns. The typical pattern vector for the "significant decrease in output power" type is [low value, normal value, low value, low value, ...], for the "abnormal equipment temperature" type it is [normal value, low value, normal value, high value, ...], and for the "abnormal communication data" type it is [normal value, normal value, low value, normal value, ...]. The current anomalous feature data segment [0.72, 0.85, 0.69, 0.48, ..., 0.65] is compared with each predefined pattern using cosine similarity calculation.

[0200] The calculation results show that the similarity to the "significantly decreased output power" pattern is 0.78, the similarity to the "abnormal equipment temperature" pattern is 0.65, and the similarity to the "abnormal communication data" pattern is 0.42. Since the highest similarity of 0.78 exceeds the threshold of 0.7, the system initially identifies the anomaly type as "significantly decreased output power." This preliminary diagnosis provides maintenance personnel with quick guidance on the problem's direction and offers crucial prior information for subsequent precise fault location and detailed diagnosis, helping to narrow down the troubleshooting scope and improve diagnostic efficiency.

[0201] Step S4: Precise fault location and detailed fault type identification based on approximate spanning tree. This step specifically includes:

[0202] Step S41: Loading System Topology and Electrical Connections. Based on the abnormal state alarm signals, abnormal feature data fragments, and preliminary abnormal type indications, a weighted graph model is constructed based on the physical topology and electrical connection information of the photovoltaic system. Specifically, the complete physical topology and detailed electrical connection information of the photovoltaic system are loaded from a preset database. This information is then cleaned and format-converted to obtain the structured topology data of the current photovoltaic system. The system loads the following information from the configuration database:

[0203] The physical location and electrical connections (series and parallel) of photovoltaic modules.

[0204] Location of the combiner box and information on the connected bus strings

[0205] Inverter location and connected combiner box information

[0206] Connection relationship of power distribution equipment

[0207] Sensor deployment location and monitored objects

[0208] The system cleans and standardizes this raw information, including removing redundant information, unifying naming rules, and converting it into a standard format (such as JSON or XML), ultimately forming structured topology data, which includes a node information table and a connection relationship table.

[0209] Step S42: Construct a weighted graph model based on the system topology. Based on the structured topology data, establish a node feature matrix, abstract each key device in the photovoltaic system as a node of the graph, and map the operating parameters of the devices as feature vectors of the nodes to obtain an initialized graph structure. Based on the initialized graph structure, transform the physical connection relationships between devices into an adjacency matrix, and transform the electrical influence relationships into association strengths to obtain the basic connection structure of the graph. Based on the abnormal feature data fragments and preliminary abnormality type indications, calculate the correlation coefficients and fault propagation probabilities between nodes in the graph structure, and use these correlation coefficients and fault propagation probabilities as edge weight attributes to obtain an edge weight matrix. Combine the node feature matrix, the basic connection structure, and the edge weight matrix to construct a weighted graph model representing the complete system topology and fault association relationships. The specific implementation is as follows:

[0210] Node feature matrix construction:

[0211] Each device node's feature vector includes: device type code (one-hot encoding), geographic location coordinates (normalized), current operating parameters (such as voltage, current, temperature, etc., normalized), historical fault frequency, etc.

[0212] The node feature matrix has an n×d dimension, where n is the number of nodes and d is the feature dimension (approximately 20-30).

[0213] Basic connection structure construction:

[0214] Adjacency matrix A: an n×n matrix, where A[i,j] = 1 indicates that there is a physical or electrical connection between nodes i and j, and 0 otherwise.

[0215] The correlation strength matrix S is an n×n matrix where S[i,j]∈[0,1] represents the influence strength of node i on node j, determined based on electrical topology relationships (such as upstream and downstream relationships, electrical distance, etc.).

[0216] Edge weight matrix calculation:

[0217] Correlation coefficient matrix C: Based on the anomaly feature data fragments, the Pearson correlation coefficients between the parameters of each node are calculated.

[0218] Fault propagation probability matrix P: Based on the initial anomaly type and historical fault propagation patterns, estimates the probability that a fault will propagate from node i to node j.

[0219] The edge weight matrix W = α·S + β·C + γ·P, where α, β, and γ are weight coefficients, satisfying α + β + γ = 1 (in this embodiment, α = 0.3, β = 0.4, and γ = 0.3).

[0220] Weighted graph model construction:

[0221] The node feature matrix, adjacency matrix, and edge weight matrix are combined into a complete graph data structure.

[0222] Use NetworkX or a similar graph processing library for storage and manipulation.

[0223] Step S43: Construct a fault-related subgraph using an approximate spanning tree algorithm. Specifically, the system applies an optimization method to compute an approximate spanning tree from an unrelated edge set. The goal of this algorithm is to quickly find an approximate minimum spanning tree or directed acyclic graph (DAG) structure in the entire system topology graph that contains nodes indicating anomalies (or nodes most relevant to the anomaly features) and has high connection weights (or indicates strong fault correlation). The algorithm steps are as follows:

[0224] initialization:

[0225] Based on the initial anomaly type indication, a set of seed nodes S (the nodes most relevant to the anomaly) is identified.

[0226] Set an empty set of edges E and a set of nodes V (initially containing the seed node).

[0227] Iterative construction:

[0228] For each node v already in V, examine all its adjacent nodes u.

[0229] Calculate the relevance score of edge (v,u): score(v,u) = W[v,u] * relevance(u), where relevance(u) is the relevance of node u to the anomaly.

[0230] Sort all edges in descending order of score, select the first k edges and add them to E, and add the corresponding new nodes to V.

[0231] If adding an edge would create a loop, then skip that edge.

[0232] Repeat the above process until V contains a preset number of nodes or the score is below a threshold.

[0233] Optimize pruning:

[0234] The constructed subgraph is optimized using a minimum spanning tree algorithm (such as Kruskal's algorithm).

[0235] Remove edges with weights below a threshold to ensure the compactness and relevance of the subgraph.

[0236] The resulting fault correlation subgraph contains a set of devices and their connections that are most relevant to the current anomaly, providing a basis for narrowing the scope and clarifying the relationships for subsequent accurate fault location.

[0237] Step S44: Fault Correlation Subgraph Analysis and Fault Root Cause Location. By analyzing the fault correlation subgraph, precise fault unit location information, detailed fault type diagnosis conclusions, and an assessment report on the fault's impact range and severity are output. Specifically, a depth-first traversal is performed on the nodes in the fault correlation subgraph to extract historical data for each node in the 10 sampling periods before the fault occurred and real-time data after the fault occurred. Parameter trend comparison analysis is performed to obtain the temporal evolution characteristics of parameter anomalies. Based on these temporal evolution characteristics, a causal relationship graph between nodes is constructed. The maximum information coefficient is used to calculate the correlation strength between nodes to obtain the critical path of fault propagation. Based on the critical path of fault propagation, combined with the electrical logic relationship of the equipment, a backtracking analysis is performed to locate the fault source equipment and obtain precise fault unit location information. The specific implementation is as follows:

[0238] Parameter trend comparison and analysis:

[0239] For each node in the subgraph, extract historical data from 50 minutes before the fault occurred (10 sampling periods) and real-time data after the fault occurred.

[0240] Calculate the rate of change, fluctuation amplitude, and trend slope of key parameters (such as voltage, current, temperature, etc.).

[0241] Identify anomalous patterns of parameter changes, such as sudden changes, gradual changes, and oscillations.

[0242] Generate the parameter anomaly time series feature vector for each node.

[0243] Causal relationship graph construction:

[0244] For each pair of nodes (i,j), calculate the maximum information coefficient (MIC) using the formula MIC(X,Y)=max_{ngrids≤B(n)}[I(X,Y) / log(ngrids)], where I(X,Y) is the mutual information and B(n) is an upper bound function of the number of grids.

[0245] MIC values ​​range from [0,1], with larger values ​​indicating stronger correlations.

[0246] Set a threshold τ = 0.7. When MIC(i,j) > τ, add a directed edge from i to j in the cause-effect graph.

[0247] The direction of the edge is determined based on temporal precedence: if the parameter change of node i precedes that of node j in time, then the direction is i→j.

[0248] Source of the fault:

[0249] In a causal graph, identify nodes with an in-degree of 0 or the minimum in-degree; these nodes are potential sources of failure.

[0250] Verification and screening are performed by combining electrical logic relationships (such as upstream and downstream relationships).

[0251] If multiple candidate sources exist, they should be ranked according to the degree of anomaly, temporal sequence, and electrical importance.

[0252] The node with the highest overall score is selected as the final source of the fault, down to the specific equipment unit (such as a photovoltaic module or an inverter module).

[0253] Step S45: Detailed Fault Type Identification and Impact Assessment. The precise fault unit location information is associated with relevant abnormal feature data fragments. A pre-set detailed fault knowledge base is queried, and pattern matching is performed using a combination of rule-based expert systems and support vector machines to obtain detailed fault type diagnostic conclusions. Based on the fault type diagnostic conclusions and fault unit location information, combined with the system topology, Monte Carlo simulation is used to assess the propagation range and impact of the fault at different time scales, resulting in an assessment report of the fault's impact range and severity. The specific implementation is as follows:

[0254] Detailed fault type identification:

[0255] The fault knowledge base includes characteristic signatures, parameter patterns, and typical cases of common faults in various photovoltaic equipment.

[0256] Rule-based expert system: Contains approximately 200 IF-THEN rules, covering common fault scenarios.

[0257] Support Vector Machine (SVM) model: Utilizing a Radial Basis Function (RBF) kernel with parameters C = 10 and gamma = 0.1, it is trained on historical fault data.

[0258] Mixed reasoning process:

[0259] First, expert system rules are applied for preliminary matching to obtain a set of candidate fault types.

[0260] Then, an SVM model is used to perform fine-grained classification of the candidate types to obtain the most likely fault types and their probabilities.

[0261] If the highest probability is below 0.8, it is marked as "fault type undetermined" and further investigation is recommended.

[0262] Fault impact assessment:

[0263] Monte Carlo simulation parameters:

[0264] Number of simulations: 1000

[0265] Time scales: Short-term (within 1 hour), Medium-term (within 24 hours), Long-term (within 7 days)

[0266] Propagation Model: Markov Chain Model Trained Based on System Topology and Historical Fault Propagation Data

[0267] Evaluation indicators:

[0268] Scope of impact: The number and extent of affected devices (e.g., local / regional / global).

[0269] Power generation loss: the expected percentage reduction in power generation.

[0270] Equipment damage risk: the probability of permanent equipment damage.

[0271] Impact on system stability: The degree of impact on the overall stable operation of the system.

[0272] Assessment report generation:

[0273] Based on the above indicators, a graded assessment result (mild / moderate / severe / critical) is generated.

[0274] Provides detailed impact descriptions and time evolution predictions

[0275] Recommended processing priorities and time limits

[0276] Detailed examples of steps S41-S45

[0277] Step S41: Loading System Topology and Electrical Connections

[0278] The topology database of this 10MW photovoltaic power plant records complete system architecture information. The power plant contains 200 strings, each string consisting of 20 315W photovoltaic modules connected in series, for a total of 4000 modules. These strings are divided into 20 groups, with 10 strings connected to a combiner box, and the 20 combiner boxes are then connected to four 2.5MW inverters. The system loads the specific connection relationships from the database: strings PV001-PV010 are connected to combiner box CB001, combiner boxes CB001-CB005 are connected to inverter INV001, and each connection point is equipped with voltage, current, and temperature sensors.

[0279] During data cleaning, the system discovered inconsistencies in some historical records, such as differences in the connection relationships of string PV055 across different documents. Through on-site verification and equipment identification confirmation, the system corrected these errors. Format conversion unified the original Excel spreadsheets and CAD drawings into structured data in JSON format, including a node attribute table (equipment ID, type, location coordinates, rated parameters) and an edge relationship table (starting node, ending node, connection type, electrical distance). The resulting structured topology data accurately reflects the actual physical and electrical connection status of the power plant.

[0280] Step S42: Construct a weighted graph model based on the system topology

[0281] The system abstracts 4,000 photovoltaic modules, 200 strings, 20 combiner boxes, 4 inverters, and related sensors and switching devices as nodes in a graph, totaling approximately 4,300 nodes. Each node's feature vector contains 25 dimensions: device type using one-hot encoding (5 dimensions), geographic location coordinates normalized to latitude and longitude (2 dimensions), current operating parameters such as voltage, current, temperature, and power (10 dimensions), and historical fault frequency, operating years, and maintenance records (8 dimensions). For example, the feature vector of a certain photovoltaic module is [0,1,0,0,0,0.523,0.687,0.85,0.76,0.92,0.78,...,0.15,2.3,0.12].

[0282] The basic connection structure is represented by an adjacency matrix with dimensions of 4300×4300. If component A and component B are connected in series in the same string, the corresponding position in the adjacency matrix is ​​1; otherwise, it is 0. The correlation strength matrix considers the degree of electrical influence: the correlation strength of series-connected components is 1.0, the correlation strength of different strings within the same combiner box is 0.8, and the correlation strength of different combiner boxes under the same inverter is 0.6. Based on the current anomaly characteristic data, the system calculates the correlation coefficient matrix between components and finds that the correlation coefficients of 10 components in a certain area are generally high (0.7-0.9), suggesting the possible existence of common influencing factors.

[0283] Step S43: Construct a fault association subgraph using an approximate spanning tree algorithm.

[0284] Based on the "significant decrease in output power" anomaly type determined in step S34, the system identifies three seed nodes: strings PV087, PV088, and PV089. The power output of these strings has decreased by 25-30% compared to normal levels. The algorithm starts from these three seed nodes and examines all their adjacent nodes. PV087 is connected to combiner box CB009, which in turn is connected to inverter INV003. CB009 is also connected to nine other strings PV081-PV090.

[0285] The system calculates the relevance score for each edge. For example, the weight of edge (PV087, CB009) is 0.85, and its relevance score is 0.85 × 0.9 = 0.765; the weight of edge (CB009, PV086) is 0.72, and its relevance score is 0.72 × 0.7 = 0.504. After iterative construction, the algorithm selects the 50 edges with the highest scores to form a fault correlation subgraph containing 67 nodes. This subgraph contains 3 abnormal strings, 2 combiner boxes, 1 inverter, and 61 related sensors and monitoring points. After optimization using the Kruskal algorithm, the final fault correlation subgraph retains 35 of the most relevant nodes and 42 high-weight connecting edges.

[0286] Step S44: Fault Association Subgraph Analysis and Fault Root Cause Location

[0287] The system performed a depth-first traversal of 35 nodes in the fault correlation subgraph, extracting historical data for each node from the 50 minutes prior to the fault occurrence. Analysis revealed that the voltage of string PV087 gradually decreased from a normal 630V to 580V within 30 minutes, the current decreased from 15.8A to 14.2A, while the module temperature increased from 45℃ to 52℃. In contrast, the parameters of the adjacent string PV086 changed very little, with the voltage only decreasing from 632V to 628V. This differentiated parameter evolution pattern provides important clues.

[0288] Through maximum information coefficient calculation, the system found that the MIC value between PV087 and its cascaded 15th component was 0.85, significantly higher than the correlation with other components (typically between 0.3 and 0.5). Further analysis of time-series data revealed that the temperature change of the 15th component led the power drop of the entire string by approximately 8 minutes, consistent with the typical propagation pattern of hot spot effects. The causal relationship diagram showed that the 15th component was the only node with an in-degree of 0, and all other anomalies could be traced back to this source. Combined with electrical logic relationship analysis, the system determined that the 15th component (device number PV087-15) in the PV087 string was the source of the fault.

[0289] Step S45: Detailed Fault Type Identification and Impact Assessment

[0290] The system correlates the fault location information (PV087-15) with anomaly feature data and queries the fault knowledge base for fault modes of individual photovoltaic modules. Rule matching by the expert system reveals that the pattern of a localized temperature increase of 8°C and a voltage decrease of 8% matches the typical characteristics of the "hot spot effect." The support vector machine model further confirms that the classification result of the input feature vector [temperature change rate: 0.8, voltage decrease magnitude: 0.08, current reduction: 0.1, power loss: 0.18] is "hot spot effect," with a confidence level of 0.92.

[0291] The impact assessment employed Monte Carlo simulation, running 1000 simulations to analyze fault propagation. Short-term (within 1 hour) impact: The probability of the component continuing to deteriorate is 78%, potentially leading to a 35-40% power reduction across the entire string, with an estimated power loss of 2.5 kWh. Medium-term (within 24 hours) impact: The hot spot may trigger bypass diode failure, affecting adjacent components and expanding the impact to 2-3 strings, resulting in a power loss of approximately 60 kWh. Long-term (within 7 days) impact: If not addressed promptly, the component may be permanently damaged, requiring replacement, with downtime losses exceeding 500 kWh. The assessment report classifies the fault severity as "moderate" and recommends addressing the issue within 24 hours to prevent further losses.

[0292] Step S5: Dynamic generation and closed-loop execution verification of intelligent self-healing strategy. This step specifically includes:

[0293] Step S51: Self-healing strategy matching and generation. Based on the fault unit location information, fault type diagnosis conclusion, and assessment report of fault impact range and severity, a self-healing operation plan is generated. Specifically, historical processing cases related to the fault type diagnosis conclusion are extracted from a preset self-healing strategy library. Similarity calculation and strategy matching are performed using a case reasoning method to obtain initial candidate self-healing operation plans. A deep reinforcement learning model is used to train a policy network based on historical operating data to generate new candidate processing plans for the current fault scenario. These new plans are then merged with the initial candidate self-healing operation plans to obtain a complete set of candidate self-healing operation plans. The specific implementation is as follows:

[0294] Case reasoning matching:

[0295] The self-healing strategy library contains historical fault handling cases, each including fault characteristics, handling methods, and effectiveness evaluations.

[0296] The top 5 cases with the highest similarity were selected, and their processing methods were extracted as initial candidate solutions.

[0297] Deep reinforcement learning models:

[0298] Network structure:

[0299] State space: A vector containing fault characteristics, system states, and environmental conditions (approximately 50 dimensions).

[0300] Action space: a set of discretized operations (combinations of approximately 30 basic operations).

[0301] Reward function: a weighted sum based on system recovery level, operational costs, and time efficiency.

[0302] Policy network: 4-layer fully connected network with hidden layer neurons of [256, 128, 64, 32] and ReLU activation function.

[0303] Value Network: A 3-layer fully connected network with 128, 64 hidden neurons and ReLU activation function.

[0304] Training parameters:

[0305] Algorithm: Proximal Policy Optimization (PPO)

[0306] Discount factor γ = 0.95

[0307] GAE parameter λ = 0.97

[0308] Learning rate = 3e-4, using Adam optimizer

[0309] Training environment: A photovoltaic system simulation environment built based on historical data.

[0310] Training rounds: 10,000 rounds, each round containing 1,000 interactive steps.

[0311] Reasoning process:

[0312] Input the current fault state vector

[0313] Policy network output action probability distribution

[0314] The top three action combinations with the highest probabilities were selected as candidate processing solutions.

[0315] Candidate solution merging:

[0316] Merge candidate solutions generated based on case-based reasoning and deep reinforcement learning.

[0317] Remove duplicate or highly similar schemes

[0318] A complete set of candidate self-healing procedures is formed (usually containing 5-8 different procedures).

[0319] Step S52: Self-Healing Solution Evaluation and Optimal Selection. A multi-objective evaluation model is constructed, using factors such as expected fault recovery time, system performance recovery degree, operational complexity, resource consumption, and potential risks as evaluation indicators. The candidate self-healing operation solution set is comprehensively scored to obtain the solution evaluation results. Based on the solution evaluation results, the weight of each evaluation indicator is determined using the analytic hierarchy process (AHP), and the comprehensive score of each solution is calculated. The solution with the highest score is selected as the optimal self-healing operation solution. The specific implementation is as follows:

[0320] Multi-objective evaluation model:

[0321] Evaluation indicators:

[0322] Fault recovery time (T): The estimated time (in minutes) required to complete the self-healing process.

[0323] Performance recovery percentage (P): The percentage of expected recovery to normal levels (0-100%)

[0324] Operational complexity (C): The number and difficulty of operation steps (1-10 points, lower is simpler).

[0325] Resource consumption (R): The human and material resources required to perform the operation (1-10 points, the lower the score, the more resources are saved).

[0326] Potential Risks (K): New problems or risks that may arise from the operation (1-10 points, the lower the score, the safer).

[0327] Scoring criteria:

[0328] Each indicator is quantified according to preset scoring rules.

[0329] For T and C, a linear mapping is used to transform them to the interval [0,1].

[0330] For P, simply divide the percentage value by 100.

[0331] For R and K, a reverse mapping is used to transform them to the [0,1] interval (the smaller the value, the higher the score).

[0332] Determining weights in the Analytic Hierarchy Process (AHP):

[0333] Construct a judgment matrix A, where A[i,j] represents the importance of indicator i relative to indicator j.

[0334] Calculate the maximum eigenvalue λ_max of the judgment matrix and its corresponding eigenvector.

[0335] The normalized eigenvectors yield the weight vector w = [w_T, w_P, w_C, w_R, w_K]

[0336] The consistency ratio CR is calculated as (λ_max - n) / (n - 1) / RI, ensuring that CR < 0.1.

[0337] The weights obtained in this embodiment are: w_T = 0.15, w_P = 0.35, w_C = 0.20, w_R = 0.10, w_K = 0.20

[0338] Overall score calculation:

[0339] For each candidate solution i, calculate its comprehensive score:

[0340] Score_i=w_T*score_T(i)+w_P*score_P(i)+w_C*score_C(i)+w_R*score_R(i)+w_K*score_K(i)

[0341] The solution with the highest score is selected as the optimal self-healing operation plan.

[0342] Step S53: Issuance and execution of self-healing operation instructions. The optimal self-healing operation scheme is decomposed into a sequence of device-level control instructions, and an instruction execution flowchart is established based on the execution sequence dependencies to obtain the operation instructions; the control module issues the operation instructions to the relevant actuators and executes the self-healing operation. The specific implementation is as follows:

[0343] Instruction breakdown:

[0344] The advanced self-healing solution is broken down into basic operational units (such as switch control, parameter adjustment, mode switching, etc.).

[0345] Assign a unique operation ID and execution priority to each operation unit.

[0346] Determine the dependencies and execution conditions between operations.

[0347] Execution flowchart construction:

[0348] Constructing Directed Acyclic Graphs (DAGs) based on operation dependencies

[0349] Determine the critical path and parallel execution branches

[0350] Set checkpoints and verification conditions to ensure the safe execution of operations.

[0351] Command issuance and execution:

[0352] Convert operation instructions into control commands that the device can recognize (such as Modbus commands, IEC61850 messages, etc.).

[0353] Commands are sent to the corresponding actuators via communication networks (such as industrial Ethernet, wireless sensor networks, etc.).

[0354] The actuator (such as a programmable logic controller (PLC), a smart circuit breaker, an inverter control unit, etc.) receives and executes instructions, and records the execution status and response result of each instruction.

[0355] Step S54: Self-Healing Effect Monitoring and Evaluation. Monitor system status parameters after the self-healing operation. Specifically, establish a multi-dimensional monitoring index system including power generation efficiency, equipment temperature, electrical parameters, and communication status. Real-time data collection is performed on key nodes during the self-healing operation to obtain a status monitoring data stream. The status monitoring data stream is compared with historical baseline data before the fault, and the recovery degree of each monitoring index is calculated. When the key index recovers to more than 90% of the baseline value, the self-healing operation is confirmed as successful, and real-time data on the execution status of the self-healing operation and the system performance recovery are obtained. The specific implementation is as follows:

[0356] Multidimensional monitoring indicator system:

[0357] Power generation efficiency indicators:

[0358] Power output ratio: Current output power / Theoretical maximum power (considering current irradiance and temperature)

[0359] Conversion efficiency: Output electrical energy / Input solar energy

[0360] Performance ratio: Actual power generation / Theoretical power generation

[0361] Equipment temperature specifications:

[0362] Component temperature distribution uniformity

[0363] Key temperature points (such as inverter IGBT temperatures)

[0364] Temperature rise rate

[0365] Electrical parameters:

[0366] Voltage stability: voltage fluctuation range and frequency

[0367] Current balance: Current deviation in each branch

[0368] Power factor

[0369] Harmonic content

[0370] Communication status indicators:

[0371] Communication success rate

[0372] Response time

[0373] Data integrity

[0374] Recovery assessment:

[0375] Baseline data selection: Historical data under the same environmental conditions before the failure occurred are selected as the baseline.

[0376] Recovery rate calculation: recovery_rate = current_value / baseline_value * 100%

[0377] Weighted composite score: weighted_recovery = ∑(w_i * recovery_rate_i), where w_i is the weight of each indicator.

[0378] Success criteria: Key indicator recovery rate ≥ 90% and weighted composite score ≥ 85%.

[0379] Step S55: Log Recording and Knowledge Base Update. A hierarchical log recording model is constructed, organizing alarm information, diagnostic processes, self-healing measures, operational responses, and performance recovery information according to chronological order and logical relationships to generate a structured event log, resulting in log records containing the entire fault diagnosis and self-healing process. Key information regarding fault characteristics, diagnostic rules, and self-healing strategies is extracted from the log records, and knowledge extraction and rule summarization are performed to establish a knowledge graph update strategy. When the self-healing operation is successful and the performance recovery exceeds expectations, the extracted knowledge and rules are updated to the fault knowledge base and self-healing strategy base to optimize the system's diagnostic and processing capabilities. The specific implementation is as follows:

[0380] Hierarchical logging model:

[0381] Event layer: Records the occurrence time, type, and basic description of key events.

[0382] Data layer: Records raw data and processing results related to the event.

[0383] Decision-making level: Records the diagnostic and decision-making processes made by the system.

[0384] Execution layer: Records detailed information and results of operation execution.

[0385] Evaluation layer: Records effectiveness evaluation and feedback information.

[0386] Knowledge extraction and rule induction:

[0387] Feature extraction: Extracting fault feature patterns and key indicators from successful cases.

[0388] Rule induction: Based on decision tree or association rule mining algorithms, new diagnostic rules are induced from successful cases.

[0389] Strategy optimization: Evaluate the effectiveness of self-healing strategies and refine efficient treatment methods.

[0390] Knowledge base update strategy:

[0391] Update trigger conditions: Self-healing operation is successful and performance recovery exceeds expectations (e.g., recovery rate > 95%).

[0392] Update process:

[0393] Extract key information from the case and form structured knowledge items.

[0394] Compare with existing knowledge base content to avoid redundancy.

[0395] Update the fault feature library, diagnostic rule library, and self-healing strategy library.

[0396] Record update logs for easy tracking and backtracking.

[0397] Detailed examples of steps S51-S55

[0398] Step S51: Self-healing strategy matching and generation

[0399] The system retrieved five historical handling cases related to the "hot spot effect" from the self-healing strategy library. Case A recorded a similar hot spot problem in PV modules, and the handling solution was to temporarily isolate the faulty string and adjust the MPPT parameters, achieving an 85% recovery rate in 45 minutes. Case B used a partial shading method to reduce the operating current of the faulty module, achieving a 70% recovery rate in 20 minutes. Using weighted Euclidean distance calculation, the current fault has a similarity of 0.83 with Case A and 0.76 with Case B.

[0400] The deep reinforcement learning model infers based on the current state vector [fault type: hotspot, location: edge component, weather: sunny, load: 75%, time: morning]. The action probability distribution output by the policy network shows that the probability of "adjusting the MPPT operating point" is 0.45, the probability of "enabling bypass protection" is 0.32, and the probability of "reducing the string operating current" is 0.23. Combining historical cases and reinforcement learning results, the system generates 6 candidate self-healing schemes: Scheme 1 - isolate the faulty string, Scheme 2 - adjust MPPT parameters, Scheme 3 - enable bypass diode, Scheme 4 - reduce operating current, Scheme 5 - local shading for cooling, and Scheme 6 - combined scheme (adjusting MPPT + enabling bypass).

[0401] Step S52: Evaluation and Optimal Selection of Self-Healing Solution

[0402] The system performs a multi-objective evaluation of six candidate solutions. The evaluation results for Solution 1 (isolieving fault chains) are as follows: fault recovery time 10 minutes (score 0.9), performance recovery rate 95% (score 0.95), operational complexity 2 points (score 0.8), resource consumption 1 point (score 0.9), and potential risk 2 points (score 0.8). The evaluation results for Solution 2 (adjusting MPPT parameters) are as follows: recovery time 15 minutes (score 0.85), performance recovery 80% (score 0.8), operational complexity 3 points (score 0.7), resource consumption 1 point (score 0.9), and potential risk 1 point (score 0.9).

[0403] Based on the weights determined by the Analytic Hierarchy Process (AHP) (recovery time 0.15, performance recovery 0.35, operational complexity 0.20, resource consumption 0.10, potential risk 0.20), the comprehensive score of each scheme is calculated. Scheme 1 score: 0.15×0.9+0.35×0.95+0.20×0.8+0.10×0.9+0.20×0.8=0.875. Scheme 2 score: 0.15×0.85+0.35×0.8+0.20×0.7+0.10×0.9+0.20×0.9=0.825. After comprehensive evaluation, Scheme 6 (combined scheme) achieves the highest score of 0.892 because it combines the precision of MPPT adjustment with the safety of bypass protection, effectively mitigating hot spot problems while maintaining high system performance.

[0404] Step S53: Issuance and execution of self-healing operation instructions

[0405] The system decomposes the optimal self-healing solution into a specific sequence of operating instructions. Step 1: Send an MPPT parameter adjustment command to inverter INV003, adjusting the operating voltage of the PV087 string from 630V to 580V, and limiting the current to 13A. Step 2: Activate the bypass diode protection function of the PV087-15 module. Step 3: Monitor the operating status of other modules to ensure that the adjustment does not affect normally functioning modules.

[0406] The operation commands are converted into standard Modbus RTU protocol messages: the command to adjust MPPT parameters is "0106001E0258XX XX" (device address 01, function code 06, register address 001E, data 0258 represents 580V). The command is sent to the inverter control unit via industrial Ethernet, and the inverter returns an acknowledgment message "0106001E 0258XX XX" after receiving the command. The bypass diode activation command is executed through the built-in controller of the intelligent component, and the entire operation takes 12 minutes. The system records the execution status of each command in real time: MPPT adjustment successful, response time 2.3 seconds; bypass activation successful, response time 1.8 seconds.

[0407] Step S54: Monitoring and Evaluation of Self-Healing Effect

[0408] Following the self-healing operation, a real-time monitoring mechanism was established. Power generation efficiency indicators showed that the power output of the PV087 string recovered from 4.2kW ​​at the time of the fault to 5.8kW, and the power output ratio increased from 67% to 92%. Equipment temperature indicators showed that the temperature of the PV087-15 module gradually decreased from a peak of 52℃ to 47℃, with a more uniform temperature distribution. Electrical parameters showed that the string voltage remained stable within the range of 580V±5V, current balance was good, and the current deviation in each branch was less than 2%.

[0409] By comparing the system with historical baseline data before the fault, the system calculated the recovery level of various indicators. Power generation efficiency recovered to 92% of the baseline value, equipment temperature recovered to 94% of the baseline value, and electrical parameter stability recovered to 96% of the baseline value. The weighted comprehensive score was 93.2%, exceeding the 90% success rate criterion. The system confirmed the successful execution of the self-healing operation and generated a detailed recovery report: the temperature of the faulty component decreased by 5°C, string power recovered by 38%, and the overall system performance impact was controlled within 2%, successfully avoiding the risk of permanent component damage.

[0410] Step S55: Log recording and knowledge base update

[0411] The system constructed a complete event log, recording the entire fault handling process in chronological order. The event layer log shows: 09:42 anomaly detection triggered, 09:45 fault location completed, 09:52 self-healing plan generated, 10:04 operation started, 10:16 operation completed, 10:30 recovery confirmed. The data layer recorded the changes in key parameters: temperature dropped from 52℃ to 47℃, and power increased from 4.2kW ​​to 5.8kW. The decision layer recorded the diagnostic reasoning process: MIC calculation → causal analysis → hotspot identification → solution evaluation → optimal selection.

[0412] The knowledge extraction process extracted new empirical patterns from successful cases: when hot spots appear on edge components, the combined strategy of "MPPT adjustment + bypass protection" is most effective, achieving a recovery rate of over 90%. The system summarized a new diagnostic rule: "IF component temperature > 50℃ AND voltage drop > 5% AND location = edge THEN hot spot probability = 0.9". Because the recovery rate of this self-healing operation (93.2%) exceeded the expected 90%, the system triggered a knowledge base update mechanism. The updates included: adding new hot spot identification patterns to the fault feature library, recording successful combined treatment schemes in the self-healing strategy library, and adding specific rules for edge component hot spots to the diagnostic rule library. These updates will improve the system's ability and efficiency in handling similar faults in the future.

[0413] like Figure 4As shown, the present invention also provides an intelligent fault diagnosis system for photovoltaic systems, comprising:

[0414] The acquisition module is used to acquire multi-source heterogeneous operating parameter data of the photovoltaic system based on the intelligent sensor network deployed at each key node of the photovoltaic system, and to perform standardized preprocessing on the multi-source heterogeneous operating parameter data to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

[0415] The calculation module is used to extract local features and calculate frequency moments of the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes, perform data summary aggregation through the frequency moment optimization algorithm, and process the data summary using a multi-modal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system.

[0416] An anomaly detection module is used to perform anomaly detection on the fused key feature vector set based on a pre-trained normal behavior pattern baseline. When an anomaly is detected, it generates an anomaly state alarm signal, related anomaly feature data fragments, and a preliminary anomaly type indication.

[0417] The fault diagnosis module is used to construct a weighted graph model based on the physical topology and electrical connection information of the photovoltaic system, according to the abnormal state alarm signal, abnormal feature data fragments and preliminary abnormal type indication, and to construct a fault association subgraph using an approximate spanning tree algorithm. By analyzing the fault association subgraph, it outputs accurate fault unit location information, detailed fault type diagnosis conclusions and an assessment report on the fault impact range and severity.

[0418] The self-healing operation module is used to generate a self-healing operation plan based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact. It then issues operation instructions to relevant actuators through the control module and executes the self-healing operation.

[0419] The working process of the present invention is described in detail below through a specific embodiment:

[0420] The intelligent fault diagnosis system for photovoltaic systems of this invention was deployed in a large-scale distributed photovoltaic power station. The power station comprises multiple photovoltaic arrays, combiner boxes, inverters, and power distribution equipment, with a total installed capacity of 10MW. The system deploys an intelligent sensor network at each key node to collect multi-source heterogeneous data such as voltage, current, temperature, and irradiance in real time.

[0421] First, the system collects multi-source heterogeneous operating parameter data of the photovoltaic system based on the intelligent sensor network deployed at each key node of the photovoltaic system. This data is then standardized and preprocessed to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream. During this process, the system cleans the raw data, identifying and processing garbled data caused by communication errors, extreme outliers exceeding ±50% of the nominal value, and sudden noise data deviating from the mean by more than three times the standard deviation for three consecutive sampling periods. Missing data is imputed: linear imputation is used for environmental parameters with changes below a preset threshold, time-series correlation-based imputation is used for periodically changing power generation parameters, and local mean imputation is used for short-term missing data. Timestamp alignment and resampling are performed, selecting 5 minutes as the baseline sampling frequency. High-frequency data is downsampled using a moving average method, and low-frequency data is upsampled using piecewise linear interpolation. Finally, unit normalization and format unification are performed, scaling data with different physical dimensions to the [0,1] interval using a min-max normalization method.

[0422] Next, the system extracts local features and calculates frequency moments from the operating parameter data stream of the multi-source heterogeneous photovoltaic system at edge nodes. It then performs data summary aggregation using a frequency moment optimization algorithm and processes the data summary using a multimodal data fusion model to obtain a fused key feature vector set containing comprehensive operating status information of the photovoltaic system. During this process, within a preset 10-minute time window, the system calculates the statistical features of the standardized operating parameter data stream of the multi-source heterogeneous photovoltaic system, updates the first and second moments using a recursive calculation method, compresses the data using a Count-MinSketch-based probabilistic data structure, and transmits local statistics between nodes using a hierarchical aggregation strategy to achieve distributed estimation of frequency moments. A multimodal data fusion model is constructed, comprising a multilayer perceptron with encoding and decoding layers, a bidirectional long short-term memory network for processing time-series data, and a three-layer convolutional neural network for feature extraction. This model is then used to process real-time data and extract low-dimensional feature representations characterizing the current overall operating status of the photovoltaic system.

[0423] At 10:00 AM on a certain day, the system performed anomaly detection on the fusion key feature vector set based on the pre-trained normal behavior pattern baseline. It detected that the reconstruction error of the fusion feature vector of a group of photovoltaic arrays exceeded the warning threshold (0.022) for three consecutive sampling cycles. The system generated an abnormal state alarm signal, recorded the relevant abnormal feature data fragments, and indicated the preliminary anomaly type as "significant decrease in output power" by matching with the predefined pattern.

[0424] The fault diagnosis module immediately constructs a weighted graph model based on the abnormal state alarm signals, abnormal feature data fragments, and preliminary abnormal type indications, using the physical topology and electrical connection information of the photovoltaic system. The system loads complete physical topology and detailed electrical connection information of the photovoltaic system from the database, establishes a node feature matrix, transforms the physical connection relationships between devices into an adjacency matrix, transforms electrical influence relationships into correlation strength, calculates the correlation coefficients and fault propagation probabilities between nodes in the graph structure, and finally constructs a weighted graph model representing the complete system topology and fault correlation relationships.

[0425] Then, the system used an approximate spanning tree algorithm to construct a fault correlation subgraph related to the anomaly. Through in-depth analysis of the fault correlation subgraph, the system performed a depth-first traversal of the nodes in the subgraph, extracting historical data for each node in the 10 sampling periods before the fault occurred and real-time data after the fault occurred, and conducted parameter trend comparison analysis; it constructed a causal relationship graph between nodes and used the maximum information coefficient to calculate the correlation strength between nodes; based on the critical path of fault propagation and combined with the electrical logic relationship of the equipment, it conducted backtracking analysis and finally located the source of the fault to a specific photovoltaic string.

[0426] Furthermore, the system correlates precise fault location information with relevant anomaly feature data fragments, queries a pre-built detailed fault knowledge base, and uses a combination of rule-based expert systems and support vector machines for pattern matching to diagnose a hot spot effect in one of the photovoltaic modules in the string. Based on the fault type diagnosis and fault location information, combined with the system topology, the system also uses Monte Carlo simulation to assess the propagation range and impact of the fault at different time scales. The assessment results show that this fault will cause a decrease in the string's power generation efficiency of approximately 30%, and if not addressed promptly, may lead to permanent damage to the module.

[0427] Based on the diagnostic results, the self-healing module extracts historical processing cases related to hot spot effects from a pre-set self-healing strategy library. It uses case-based reasoning to calculate similarity and match strategies, while simultaneously generating new candidate processing schemes using a deep reinforcement learning model. Ultimately, two main candidate self-healing schemes were obtained: 1) temporarily isolating the fault cluster; 2) adjusting MPPT parameters to mitigate the impact of hot spots. The system constructs a multi-objective evaluation model, using factors such as expected fault recovery time, system performance recovery degree, operational complexity, resource consumption, and potential risks as evaluation indicators to comprehensively score the two schemes. After determining the weights of each evaluation indicator and calculating the comprehensive score using the analytic hierarchy process (AHP), the system selects scheme 2 as the optimal self-healing scheme and decomposes it into specific control commands, which are then issued to the relevant inverters through the control module.

[0428] After the self-healing operation was performed, the system established a multi-dimensional monitoring index system including power generation efficiency, equipment temperature, electrical parameters, and communication status, and collected data in real time at key nodes during the self-healing process. By comparing the monitoring data with historical baseline data before the fault occurred, the system found that the string temperature distribution gradually became uniform, and the output power recovered to 85% of the normal level. Although it did not reach the 90% complete recovery standard, considering the characteristics of the hot spot effect, this degree of recovery is already a good result.

[0429] Finally, the system constructs a hierarchical log recording model, organizing alarm information, diagnostic processes, self-healing measures, operational responses, and performance recovery information throughout the entire fault diagnosis and self-healing process according to chronological order and logical relationships, generating a structured event log. The system extracts key information about fault characteristics, diagnostic rules, and self-healing strategies from the log records, performs knowledge extraction and rule summarization, and updates this information to the fault knowledge base and self-healing strategy base. Simultaneously, it generates maintenance suggestions, reminding operations and maintenance personnel to check and clean up the component during the next routine maintenance.

[0430] As can be seen from this embodiment, the present invention can quickly and accurately locate faults in photovoltaic systems and automatically perform appropriate self-healing operations, which greatly improves the reliability and operating efficiency of the system and reduces the need for manual intervention.

[0431] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0432] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of the intelligent fault diagnosis method for a photovoltaic system described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0433] Furthermore, this disclosure also provides a computer program product storing a computer program. When the computer program is run by a processor, it executes the steps of a photovoltaic system intelligent fault diagnosis method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0434] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0435] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0436] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0437] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0438] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0439] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for intelligent fault diagnosis of a photovoltaic system, characterized in that, include: Based on the intelligent sensor network deployed at each key node of the photovoltaic system, multi-source heterogeneous operating parameter data of the photovoltaic system are collected, and the multi-source heterogeneous operating parameter data is standardized and preprocessed to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream. Local features are extracted and frequency moments are calculated from the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes. Data summary aggregation is performed through the frequency moment optimization algorithm. The data summary is processed using a multimodal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system. Anomaly detection is performed on the fused key feature vector set based on a pre-trained baseline of normal behavior patterns. When an anomaly is detected, an anomaly state alarm signal, related anomaly feature data fragments, and preliminary anomaly type indication are generated. Based on the abnormal state alarm signals, abnormal feature data fragments, and preliminary abnormal type indications, a weighted graph model is constructed based on the physical topology and electrical connection information of the photovoltaic system. An approximate spanning tree algorithm is used to construct a fault association subgraph. By analyzing the fault association subgraph, accurate fault unit location information, detailed fault type diagnosis conclusions, and an assessment report on the fault impact range and severity are output. Based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact, a self-healing operation plan is generated. The control module then issues operation instructions to the relevant actuators and executes the self-healing operation.

2. The method according to claim 1, characterized in that, The standardized preprocessing of the multi-source heterogeneous operating parameter data includes: Data cleaning is performed on the multi-source heterogeneous operating parameter data to identify and process garbled data caused by communication errors, extreme outliers exceeding the nominal value by ±50%, and sudden noise data that deviates from the mean by more than 3 times the standard deviation for three consecutive sampling periods, resulting in a pre-cleaned raw dataset. The original dataset that has been preliminarily cleaned is checked for missing data. Linear interpolation is used for environmental parameters whose changes are below a preset threshold. Interpolation based on time-series correlation is used for periodically changing power generation parameters. Local mean interpolation is used for short-term missing data to obtain a complete original dataset. The original dataset is timestamped and resampled. A 5-minute sampling frequency is selected as the baseline sampling frequency. High-frequency data is downsampled using a moving average method, and low-frequency data is upsampled using piecewise linear interpolation to obtain a time-aligned multi-source dataset. The time-aligned multi-source dataset is subjected to unit normalization and format unification operations. Data with different physical dimensions are scaled to the [0,1] interval using the min-max normalization method and converted into a unified data structure to obtain the standardized multi-source heterogeneous photovoltaic system operating parameter data stream.

3. The method according to claim 1, characterized in that, Local feature extraction and frequency moment calculation of the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes include: Within a preset 10-minute time window, the statistical characteristics of the standardized multi-source heterogeneous photovoltaic system operating parameter data stream are calculated. The first and second moments are updated using a recursive calculation method to obtain a distributed dataset containing local features and their frequency moments. The statistical characteristics include mean, standard deviation, peak factor, skewness, and kurtosis. For the distributed dataset containing local features and their frequency moments, a probabilistic data structure based on Count-Min Sketch is applied for data compression. A hierarchical aggregation strategy is used to transfer local statistics between nodes to achieve distributed estimation of frequency moments, thereby obtaining an optimized frequency moment summary of the global data distribution.

4. The method according to claim 3, characterized in that, The data summary is processed using a multimodal data fusion model, including: Based on the distributed dataset containing local features and the optimized frequency moment summary of the global data distribution, a multimodal data fusion model is constructed, which includes a multilayer perceptron with an encoding layer and a decoding layer, a bidirectional long short-term memory network for processing time-series data, and a three-layer convolutional neural network for feature extraction. The multimodal data fusion model is trained using a cross-validation method to obtain a trained multimodal data fusion model. Using the trained multimodal data fusion model, the standardized multi-source heterogeneous photovoltaic system operation parameter data stream that is input in real time is processed to extract low-dimensional feature representations that characterize the current overall operation status of the photovoltaic system, thereby obtaining the fusion key feature vector set containing comprehensive operation status information of the photovoltaic system.

5. The method according to claim 1, characterized in that, Anomaly detection is performed on the fused key feature vector set based on a pre-trained baseline of normal behavior patterns, including: Based on the historical fusion key feature vector set during the normal operation of the photovoltaic system, an autoencoder model containing three encoding layers and three decoding layers is constructed. By minimizing the reconstruction error, a compact representation of the normal data is learned, and a baseline model representing the normal operation state of the photovoltaic system is obtained. Each feature vector in the fusion key feature vector set is compared with the baseline model to calculate the magnitude of the reconstruction error and obtain the quantized real-time state deviation value. Based on the statistical distribution of historical operating data, a 95% confidence interval is set as the early warning threshold and a 99% confidence interval is set as the alarm threshold. When the deviation value exceeds the early warning threshold for three consecutive sampling periods or exceeds the alarm threshold once, the abnormal state alarm signal and related abnormal feature data fragments are generated. Based on the abnormal state alarm signal and related abnormal feature data fragments, the pattern of the abnormal features is matched with a predefined macroscopic abnormal category to obtain the preliminary abnormal type indication.

6. The method according to claim 1, characterized in that, A weighted graph model is constructed based on the physical topology and electrical connection information of the photovoltaic system, including: The complete physical topology information and detailed electrical connection information of the photovoltaic system are loaded from the preset database. The information is then cleaned and converted to obtain the structured topology data of the current photovoltaic system. Based on the structured topology data, a node feature matrix is ​​established, and each key device in the photovoltaic system is abstracted as a node of the graph. The operating parameters of the devices are mapped as feature vectors of the nodes to obtain the initialized graph structure. Based on the initialized graph structure, the physical connection relationships between devices are transformed into an adjacency matrix, and the electrical influence relationships are transformed into association strengths, thus obtaining the basic connection structure of the graph; Based on the abnormal feature data fragments and preliminary abnormal type indications, the correlation coefficients and fault propagation probabilities between nodes in the graph structure are calculated, and the correlation coefficients and fault propagation probabilities between nodes in the graph structure are used as the weight attributes of the edges to obtain the edge weight matrix. By combining the node feature matrix, basic connection structure, and edge weight matrix, a weighted graph model representing the complete topology and fault correlation of the system is constructed.

7. The method according to claim 1, characterized in that, By analyzing the fault correlation subgraph, the system outputs precise fault unit location information, detailed fault type diagnostic conclusions, and an assessment report on the fault's impact range and severity, including: A depth-first traversal is performed on the nodes in the fault association subgraph to extract the historical data of each node in the 10 sampling periods before the fault occurred and the real-time data after the fault occurred. Parameter trend comparison analysis is performed to obtain the time-series evolution characteristics of parameter anomalies. Based on the temporal evolution characteristics of the aforementioned parameter anomalies, a causal relationship graph between nodes is constructed, and the correlation strength between nodes is calculated using the maximum information coefficient to obtain the critical path of fault propagation. Based on the critical path of fault propagation, and combined with the electrical logic relationship of the equipment, a backtracking analysis is performed to locate the fault source equipment and obtain precise fault unit location information. The precise fault unit location information is associated with relevant abnormal feature data fragments, a pre-set detailed fault knowledge base is queried, and a pattern matching method combining rule-based expert system and support vector machine is used to obtain detailed fault type diagnosis conclusions. Based on the fault type diagnosis conclusions and fault unit location information, combined with the system topology, the Monte Carlo simulation method is used to evaluate the propagation range and impact of the fault at different time scales, and an assessment report on the impact range and severity of the fault is obtained.

8. The method according to claim 1, characterized in that, Generate a self-healing operation plan, including: Historical processing cases related to the fault type diagnosis conclusion are extracted from the preset self-healing strategy library, and similarity calculation and strategy matching are performed using the case reasoning method to obtain the initial candidate self-healing operation scheme. Using a deep reinforcement learning model, a policy network is trained based on historical operating data to generate new candidate processing schemes for the current fault scenario. These schemes are then merged with the initial candidate self-healing operation schemes to obtain a complete set of candidate self-healing operation schemes. A multi-objective evaluation model is constructed, which uses factors such as expected fault recovery time, system performance recovery degree, operational complexity, resource consumption and potential risks as evaluation indicators to comprehensively score the candidate self-healing operation scheme set and obtain the scheme evaluation results. Based on the evaluation results of the proposed schemes, the weight of each evaluation index is determined using the analytic hierarchy process (AHP), the comprehensive score of each scheme is calculated, and the scheme with the highest score is selected as the optimal self-healing operation scheme. The optimal self-healing operation scheme is decomposed into a sequence of control instructions at the device level, and an instruction execution flowchart is established based on the execution sequence dependency to obtain the operation instructions.

9. The method according to claim 1, characterized in that, The method further includes: monitoring system status parameters after the self-healing operation, including: A multi-dimensional monitoring index system including power generation efficiency, equipment temperature, electrical parameters and communication status is established, and real-time data is collected on key nodes in the self-healing operation process to obtain the status monitoring data stream of the execution process; The status monitoring data stream is compared with the historical baseline data before the failure, and the recovery degree of each monitoring indicator is calculated. When the key indicators recover to more than 90% of the baseline value, the self-healing operation is confirmed to be successful, and real-time data on the execution status of the self-healing operation and the system performance recovery status are obtained. A hierarchical log recording model is constructed, and information such as alarm information, diagnosis process, self-healing measures, operation response and performance recovery are organized according to time sequence and logical relationship to generate a structured event log, thus obtaining the log record containing the entire fault diagnosis and self-healing process. Extract key information on fault characteristics, diagnostic rules, and self-healing strategies from the log records, perform knowledge extraction and rule summarization, and establish a knowledge graph update strategy. When the self-healing operation is successful and the performance recovery exceeds expectations, the extracted knowledge and rules are updated to the fault knowledge base and self-healing strategy base to optimize the system's diagnostic and processing capabilities.

10. A photovoltaic system intelligent fault diagnosis system, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous operating parameter data of the photovoltaic system based on the intelligent sensor network deployed at each key node of the photovoltaic system, and to perform standardized preprocessing on the multi-source heterogeneous operating parameter data to obtain a standardized multi-source heterogeneous photovoltaic system operating parameter data stream. The calculation module is used to extract local features and calculate frequency moments of the operating parameter data stream of the multi-source heterogeneous photovoltaic system at the edge nodes, perform data summary aggregation through the frequency moment optimization algorithm, and process the data summary using a multi-modal data fusion model to obtain a fusion key feature vector set containing comprehensive operating status information of the photovoltaic system. An anomaly detection module is used to perform anomaly detection on the fused key feature vector set based on a pre-trained normal behavior pattern baseline. When an anomaly is detected, it generates an anomaly state alarm signal, related anomaly feature data fragments, and a preliminary anomaly type indication. The fault diagnosis module is used to construct a weighted graph model based on the physical topology and electrical connection information of the photovoltaic system, according to the abnormal state alarm signal, abnormal feature data fragments and preliminary abnormal type indication, and to construct a fault association subgraph using an approximate spanning tree algorithm. By analyzing the fault association subgraph, it outputs accurate fault unit location information, detailed fault type diagnosis conclusions and an assessment report on the fault impact range and severity. The self-healing operation module is used to generate a self-healing operation plan based on the fault unit location information, fault type diagnosis conclusion, and assessment report on the scope and severity of fault impact. It then issues operation instructions to relevant actuators through the control module and executes the self-healing operation.

Images