Fault identification method and system for photovoltaic system

Abstract

The invention relates to the technical field of photovoltaic systems, and particularly discloses a fault identification method and system for a photovoltaic system, and the method comprises the steps: collecting data in real time through environment, electrical parameters and an equipment state monitoring sensor, carrying out the cleaning and standardization, extracting time domain, frequency domain and time frequency features, and screening a feature subset through a correlation analysis and feature importance sorting algorithm; detecting and classifying faults by using a hybrid model composed of an isolated forest algorithm and a random forest classifier; positioning a fault subsystem and analyzing a root cause by means of a hierarchical diagnosis strategy, a graph neural network and a Bayesian reasoning algorithm; and periodically updating the model based on the new data. The method can quickly and accurately identify and position faults, adapts to a complex environment, reduces the operation and maintenance cost, and improves the operation reliability and operation and maintenance efficiency of a photovoltaic system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic systems, and particularly to a fault identification method and system for a photovoltaic system. BACKGROUND

[0002] In actual operation, photovoltaic systems face many challenges, among which fault problems seriously affect the power generation efficiency and reliability of the system. Photovoltaic systems are complex in composition, covering photovoltaic components, inverters, combiner boxes, monitoring systems and many other devices, and the operating environment is complex and changeable, being exposed to the outdoors for a long time, and being easily affected by natural factors such as light intensity fluctuations, extreme temperatures (high temperature, low temperature), wind and sand, rain and snow, humidity changes, etc. Strong light will cause the temperature of the photovoltaic component to rise, thereby causing power attenuation; low temperature in winter may cause the performance of the internal electronic components of the inverter to decrease, and even faults to occur; wind and sand weather will wear the surface of the photovoltaic component, reduce the light transmittance, and affect the power generation efficiency. In addition, factors such as aging of the equipment itself, manufacturing defects, improper installation, etc. will also cause various faults, such as hidden cracks and hot spots of photovoltaic components, damage to power modules and communication faults of inverters, and poor line contact of combiner boxes, etc.

[0003] At present, the traditional fault identification methods for photovoltaic systems mainly include a threshold value judgment-based method, a simple electrical parameter analysis method and a manual inspection method. The threshold value judgment-based method is to pre-set the normal range threshold of key electrical parameters (such as voltage, current and power), and when the real-time monitoring data exceeds the range, it is determined that there may be a fault. However, this method has obvious defects. On the one hand, since the output power of the photovoltaic system is greatly affected by environmental factors such as light intensity and temperature, for example, in the morning and evening when the light intensity is weak, the output voltage and current of the photovoltaic component are naturally low, which is easy to cause false alarms; on the other hand, this method can only detect faults with obviously abnormal parameters, and cannot timely detect early potential faults and hidden faults (such as small hidden cracks in photovoltaic components which have little effect on electrical parameters in the early stage).

[0004] The simple electrical parameter analysis method usually only analyzes a few parameters such as voltage and current, and lacks deep mining and comprehensive analysis of data. For example, only by observing the stability of the output voltage and current of the inverter to determine whether the system is normal, the mutual relationship between the internal devices of the system and the fault propagation path cannot be comprehensively understood, and when multiple faults occur concurrently or a fault spreads from one device to other devices, it is difficult to accurately determine the root cause of the fault.

[0005] The artificial inspection method relies on the operation and maintenance personnel to regularly go to the scene to check, test and record the equipment of the photovoltaic system. This way not only consumes a lot of manpower, material resources and time cost, but also has a long inspection cycle and cannot realize real-time monitoring. In some large photovoltaic power stations, due to the wide occupation area and large number of equipment, artificial inspection is difficult to be comprehensive and detailed, and it is easy to miss inspection. At the same time, artificial inspection is greatly affected by the professional level and work experience of the operation and maintenance personnel, and there are differences in the judgment standard and identification ability of different personnel to faults, so that the accuracy and consistency of fault identification cannot be guaranteed.

[0006] With the continuous expansion of the photovoltaic system and the increasing demand for intelligent development, the traditional fault identification method has low detection sensitivity, inaccurate fault positioning, cannot adapt to complex and changeable operating environment, and has high operation and maintenance cost, which has become increasingly prominent and cannot meet the requirements of efficient and stable operation of modern photovoltaic systems.

[0007] Therefore, it has become an urgent problem for people to develop a photovoltaic system fault identification method and system that can quickly and accurately identify and locate faults, adapt to complex environments and reduce operation and maintenance costs. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a photovoltaic system fault identification method and system, which realizes rapid and accurate identification and positioning of photovoltaic system faults through multi-source data acquisition and processing, advanced feature extraction and screening technology, hybrid fault detection model and hierarchical diagnosis strategy, and improves the operation reliability and operation efficiency of the photovoltaic system.

[0009] To solve the above technical problems, the technical scheme provided by the present application is as follows: a photovoltaic system fault identification method, comprising the following steps:

[0010] S1, through the environmental sensors, electrical parameter sensors and equipment state monitoring sensors distributed in the photovoltaic system, real-time acquisition of environmental data, electrical parameters and equipment state data, and cleaning and standardization processing of the collected raw data;

[0011] S2, based on the preprocessed data, extracting time domain features, frequency domain features and time-frequency features, and using correlation analysis and feature importance sorting algorithm to screen out the most representative feature subset for fault identification;

[0012] S3, inputting the feature subset into the pre-trained hybrid fault detection model, detecting abnormal behavior based on the unsupervised learning model and generating preliminary fault alarm, and then verifying and classifying the alarm through the supervised learning model to determine the fault type;

[0013] S4, adopt hierarchical diagnosis strategy to locate the subsystem where the fault is located, use graph neural network to construct photovoltaic system topology model, combine Bayesian inference algorithm to analyze fault propagation path, determine the root cause of the fault, and generate a diagnosis report containing fault location, type, severity and repair suggestion;

[0014] S5, based on the newly emerged fault cases and system operation data, regularly update the fault feature library and detection model, use transfer learning and incremental learning technology to improve the generalization ability of the model.

[0015] Further, the specific screening method of the feature subset is as follows:

[0016] Calculate the Pearson correlation coefficient matrix between the time domain features, frequency domain features and time-frequency features, identify the feature pairs with absolute correlation coefficient exceeding the preset threshold, for each pair of highly correlated features, calculate the correlation with the fault label, keep the features with higher correlation with the fault label, and remove the redundant features;

[0017] Use random forest algorithm to score the importance of the features screened by correlation analysis, and sort the features in descending order based on the feature importance score;

[0018] Select a preset number of top important features from the sorted features to form the final feature subset.

[0019] Further, the unsupervised learning algorithm uses the Isolation Forest algorithm, and the supervised learning algorithm uses the Random Forest classifier.

[0020] Further, in step S3, the specific method for determining the fault type is as follows:

[0021] Input the feature subset into the Isolation Forest algorithm, calculate the anomaly score of each sample, dynamically set the anomaly threshold based on the statistical characteristics of historical data, and generate a preliminary fault alarm when the sample anomaly score exceeds the threshold;

[0022] Input the sample features that trigger the alarm into the random forest classifier, which is trained by multiple fault samples, and output the fault type probability distribution;

[0023] If the highest probability category score exceeds the preset confidence threshold, it is confirmed as the corresponding fault type; if it does not exceed the threshold, start the fuzzy decision mechanism, and make a second judgment combining the anomaly score, environmental parameters and equipment running state;

[0024] Based on real-time environmental parameters and system running state, dynamically correct the anomaly threshold of Isolation Forest through Gaussian process regression, and reduce the false positive rate caused by environmental fluctuations.

[0025] Further, in step S4, the specific method for determining the root cause of the fault is as follows:

[0026] Based on the fault type output by the hybrid fault detection model, a preset fault-subsystem mapping matrix is matched, and a candidate subsystem set possibly containing faults is preliminarily determined;

[0027] For photovoltaic strings in the candidate subsystem, an electrical parameter space-time correlation network is constructed, the running state similarity between the strings is calculated, and an abnormal string cluster deviating from the normal mode is identified;

[0028] For the abnormal string cluster, device-level time sequence features are extracted, and a trained graph neural network model is input, wherein the graph neural network model is based on the node-edge relationship of the photovoltaic system topology, and the fault propagation path is inferred through the message passing mechanism;

[0029] The fault probability distribution output by the graph neural network model, the historical fault rate data of the device and the real-time environmental factors are integrated, and the posterior fault probability of each device is calculated by Bayesian inference;

[0030] For the device with a posterior probability exceeding a confidence threshold, a neighboring device state consistency check is started, and if there is a cascading anomaly, the device is confirmed as a fault source, otherwise the diagnosis path is adjusted backtracking;

[0031] A topology thermodynamic diagram containing a fault propagation path, an impact range and a root cause probability is generated, and key device nodes for recommended inspection are labeled.

[0032] Further, the environmental data includes light intensity, temperature, humidity and wind speed, the electrical parameters include voltage, current and power, and the device state data includes inverter working state and component temperature difference.

[0033] Further, the time domain features include mean, variance and peak value, the frequency domain features include power spectral density and frequency component, and the time-frequency features include wavelet transform coefficients.

[0034] The application also provides a photovoltaic system fault identification system for implementing the above method, comprising:

[0035] A data acquisition module comprising environmental sensors, electrical parameter sensors and device state monitoring sensors, for real-time acquisition of environmental data, electrical parameters and device state data, and transmission of the data to a data processing module;

[0036] A data processing module for cleaning and standardizing the collected raw data, extracting time domain features, frequency domain features and time-frequency features, and performing feature selection to obtain a feature subset;

[0037] The fault detection and classification module includes a pre-trained isolation forest algorithm and a random forest classifier, which is used to receive a feature subset, perform anomaly detection, preliminary fault alarm generation, fault type verification and classification;

[0038] The fault location and root cause analysis module uses a graph neural network and a Bayesian inference algorithm to perform fault location and root cause analysis based on the fault type, and generates a diagnostic report.

[0039] The model updating module regularly updates the fault feature library and detection model based on new fault cases and system operation data, and uses transfer learning and incremental learning techniques to optimize model performance.

[0040] Compared with the prior art, the present application has the following advantages:

[0041] The present application collects environmental, electrical and equipment state data through multiple sensors, and performs cleaning, standardization and feature extraction screening, fully utilizes multi-source information, improves data quality and fault feature representativeness, and lays a foundation for accurate fault identification.

[0042] The present application uses a hybrid fault detection model combining isolation forest algorithm and random forest classifier, the isolation forest algorithm can quickly detect abnormal behavior, the random forest classifier can accurately classify fault types, the two complement each other, and improve the accuracy and efficiency of fault detection.

[0043] The present application establishes a dynamic threshold mechanism, adjusts the fault detection sensitivity based on historical data and real-time environmental parameters, dynamically corrects the abnormal threshold value through Gaussian process regression, and effectively reduces the false alarm rate caused by environmental fluctuations.

[0044] The present application uses hierarchical diagnosis strategy, graph neural network and Bayesian inference algorithm, can accurately determine the subsystem where the fault occurs, deeply analyzes the fault propagation path and root cause, and generates a detailed diagnostic report to provide accurate guidance for maintenance.

[0045] The present application regularly updates the fault feature library and detection model based on new data, combines transfer learning and incremental learning techniques, so that the model can continuously adapt to new fault conditions and operating environment, and improve the generalization ability and long-term reliability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0046] Fig. 1 is a flowchart of a fault identification method of a photovoltaic system of the present application.

[0047] Fig. 2 is a system block diagram of a fault identification system of a photovoltaic system of the present application. DETAILED DESCRIPTION

[0048] Various exemplary embodiments of the present application will be described herein below with reference to the accompanying drawings. Note that the relative arrangement, numerical expressions, and numerical values of components and steps set forth in these embodiments are not limiting to the scope of the present application unless otherwise specifically stated.

[0049] The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the scope of the application its application or uses.

[0050] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered as part of the specification.

[0051] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary, and not as a limitation. Thus, other examples of the exemplary embodiments can have different values.

[0052] A fault identification method and system for a photovoltaic system will be described in further detail below with reference to the accompanying drawings.

[0053] In conjunction with the accompanying Figs. 1-2 A detailed description of the present application will be given.

[0054] A fault identification method for a photovoltaic system, specifically comprising the following steps:

[0055] Data acquisition and preprocessing: through the environmental sensors, electrical parameter sensors and equipment state monitoring sensors distributed in the photovoltaic system, real-time acquisition of environmental data, electrical parameters and equipment state data, and cleaning and standardization processing of the collected raw data. Among them, the environmental data includes light intensity, temperature, humidity and wind speed; the electrical parameters include voltage, current and power; the equipment state data includes inverter working state and component temperature difference.

[0056] Feature extraction and screening: based on the preprocessed data, time domain features, frequency domain features and time-frequency features are extracted, and correlation analysis and feature importance sorting algorithm are used to screen out the most representative feature subset for fault identification. The time domain features include mean, variance and peak value; the frequency domain features include power spectral density and frequency component; the time-frequency features include wavelet transform coefficients. The specific screening method of the feature subset is as follows:

[0057] Calculate the Pearson correlation coefficient matrix between the time domain features, frequency domain features and time-frequency features, identify the feature pairs with absolute correlation coefficient exceeding the preset threshold, for each pair of highly correlated features, calculate the correlation with the fault label, retain the features with higher correlation with the fault label, and remove the redundant features.

[0058] The features screened by the correlation analysis are scored for importance using a random forest algorithm, and the features are sorted in descending order based on the feature importance scores.

[0059] The top important features in the sorted features are selected in a preset number to form a final feature subset.

[0060] Fault detection and classification: The feature subset is input into a pre-trained hybrid fault detection model, abnormal behavior is detected based on an unsupervised learning model, and a preliminary fault alarm is generated, and the alarm is verified and classified by a supervised learning model to determine the fault type. The unsupervised learning algorithm uses an isolation forest algorithm, and the supervised learning algorithm uses a random forest classifier. The specific method for determining the fault type is as follows:

[0061] The feature subset is input into the isolation forest algorithm, the anomaly score of each sample is calculated, and the anomaly threshold is dynamically set based on the statistical characteristics of the historical data. When the sample anomaly score exceeds the threshold, a preliminary fault alarm is generated.

[0062] The sample features that trigger the alarm are input into the random forest classifier, which is trained on multi-class fault samples, and outputs a probability distribution of the fault type.

[0063] If the highest probability category score exceeds the pre-set confidence threshold, it is confirmed as the corresponding fault type; if it does not exceed the threshold, a fuzzy decision mechanism is started, and a secondary judgment is made in combination with the anomaly score, environmental parameters and device operating state.

[0064] Based on real-time environmental parameters and system operating state, the anomaly threshold of the isolation forest is dynamically corrected by Gaussian process regression to reduce the false alarm rate caused by environmental fluctuations.

[0065] Fault location and root cause analysis: A hierarchical diagnosis strategy is used to locate the subsystem where the fault occurs, a graph neural network is used to construct a photovoltaic system topology model, a Bayesian inference algorithm is used to analyze the fault propagation path, and the root cause of the fault is determined, and a diagnosis report containing the fault location, type, severity and repair suggestions is generated. The specific method for determining the root cause of the fault is as follows:

[0066] Based on the fault type output by the hybrid fault detection model, a pre-set fault-subsystem mapping matrix is matched to preliminarily determine a candidate subsystem set that may contain the fault.

[0067] For photovoltaic strings in the candidate subsystem, a spatio-temporal correlation network of electrical parameters is constructed, the operating state similarity between the strings is calculated, and an abnormal string cluster that deviates from the normal mode is identified.

[0068] For abnormal group string clusters, extract device-level timing features, input trained graph neural network model, model based on photovoltaic system topology structure to construct node-edge relationship, and infer fault propagation path through message passing mechanism.

[0069] Integrate the fault probability distribution output by the graph neural network, the historical fault rate data of the device, and the real-time environmental factors, and use Bayesian inference to calculate the posterior fault probability of each device.

[0070] For devices with posterior probability exceeding the confidence threshold, start adjacent device state consistency check, if there is cascading anomaly, confirm as fault source, otherwise backtrack and adjust the diagnosis path.

[0071] Generate a topology heat map containing fault propagation path, impact range and root cause probability, and label key device nodes for recommended inspection.

[0072] Model updating and optimization: based on new fault cases and system operation data, regularly update fault feature library and detection model, use transfer learning and incremental learning techniques to improve model generalization ability. In the fault detection and early warning step, based on historical data and real-time environmental parameters, establish a dynamic threshold mechanism to adaptively adjust the fault detection sensitivity.

[0073] Based on the above fault identification method, the application also provides a photovoltaic system fault identification system, which comprises:

[0074] Data acquisition module: composed of environmental sensors, electrical parameter sensors and device state monitoring sensors, used for real-time acquisition of environmental data, electrical parameters and device state data, and data transmission to the data processing module.

[0075] Data processing module: clean and standardize the collected raw data, extract time domain features, frequency domain features and time-frequency features, and perform feature selection to obtain feature subsets.

[0076] Fault detection and classification module: contains pre-trained isolation forest algorithm and random forest classifier, used to receive feature subsets, perform anomaly detection, preliminary fault alarm generation, fault type verification and classification.

[0077] Fault location and root cause analysis module: uses graph neural network and Bayesian inference algorithm to perform fault location and root cause analysis based on fault type, and generates a diagnosis report.

[0078] Model updating module: based on new fault cases and system operation data, regularly update fault feature library and detection model, use transfer learning and incremental learning techniques to optimize model performance.

[0079] The specific implementation process of the photovoltaic system fault identification method and system of the application is as follows:

[0080] I. Data Collection and Preprocessing

[0081] In a large photovoltaic power station, light intensity sensors, temperature sensors, humidity sensors, and wind speed sensors are deployed to collect environmental data; voltage sensors, current sensors, and power sensors are used to collect electrical parameters; inverter state monitoring sensors and component temperature sensors are used to collect equipment state data. The sensors collect data at a frequency of 1 per second and transmit the data to the data processing center. The data processing center cleans the collected raw data, removes outliers and missing values, and then uses a normalization method to standardize the data, mapping the data to the [0, 1] interval, to facilitate subsequent feature extraction.

[0082] II. Feature Extraction and Selection

[0083] For the preprocessed data, time domain features (such as mean, variance, peak value), frequency domain features (such as power spectral density, frequency components), and time-frequency features (such as wavelet transform coefficients) are extracted. The Pearson correlation coefficient matrix between each feature is calculated, and the absolute value threshold of the correlation coefficient is set to 0.8 to identify highly correlated feature pairs. For example, if the correlation coefficient of two frequency domain features is 0.9, calculate their correlation with the fault label, retain the feature with higher correlation, and remove redundant features. Then use the random forest algorithm to score the importance of the remaining features, sort them from high to low, and select the top 20 important features as the final feature subset.

[0084] III. Fault Detection and Classification

[0085] The feature subset is input into the Isolation Forest algorithm, and the abnormal threshold is set to 3 times the standard deviation based on historical data statistics. When the anomaly score of a sample exceeds the threshold, a preliminary fault alarm is generated. The sample features that triggered the alarm are input into the random forest classifier, which is pre-trained using sample data containing 10 different fault types. If the highest probability category score output by the random forest classifier exceeds the confidence threshold of 0.7, it is confirmed as the corresponding fault type; if it does not exceed, a fuzzy decision mechanism is used for secondary judgment based on the anomaly score, current environmental parameters (such as light intensity, temperature), and device operating state (such as inverter output power). At the same time, based on real-time environmental parameters and system operating state, the Gaussian process regression is used to dynamically correct the abnormal threshold of the Isolation Forest, reducing the false alarm rate.

[0086] IV. Fault Location and Root Cause Analysis

[0087] When the hybrid fault detection model determines the fault type, it matches the preset fault-subsystem mapping matrix to preliminarily determine the candidate subsystems that may have faults. For the photovoltaic strings in the candidate subsystems, a spatiotemporal correlation network of electrical parameters is constructed, the operating state similarity between the strings is calculated, and an abnormal string cluster deviating from the normal mode is identified. The time sequence features of the devices in the abnormal string cluster are extracted and input into a graph neural network model constructed based on the topology of the photovoltaic system. The fault propagation path is inferred through a message passing mechanism. The fault probability distribution output by the graph neural network, the historical fault rate data of the devices, and the real-time environmental factors are integrated, and the posterior fault probability of each device is calculated using Bayesian inference. When the posterior probability of a device exceeds a confidence threshold of 0.8, a consistency check of the adjacent device states is started, and if a cascading anomaly is found in the adjacent devices, the device is confirmed as the fault source. A topology heat map containing the fault propagation path, the impact range, and the root cause probability is generated, and key device nodes for inspection are labeled.

[0088] V. Model updating and optimization

[0089] New fault cases and system operation data are collected every week and added to the fault feature library. Migration learning and incremental learning techniques are used to update the fault detection model. For example, when a new fault type occurs, the knowledge of the existing model is transferred to the detection of the new fault type using migration learning, and the model is further trained using new data through incremental learning to improve the model's ability to identify new fault types. At the same time, the dynamic threshold mechanism is continuously optimized based on historical data and real-time environmental parameters to keep the fault detection sensitivity at an optimal level.

[0090] The above describes the present application and its embodiments, which are not limiting, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In summary, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structural methods and embodiments of the technical solutions should belong to the protection scope of the present application.

Claims

1. A photovoltaic system fault identification method, characterized in that: The following steps are involved: S1. Collect environmental data, electrical parameter data, and equipment status data in real time through environmental sensors, electrical parameter sensors, and equipment status monitoring sensors distributed throughout the photovoltaic system, and clean and standardize the collected raw data. S2. Based on the preprocessed data, extract time domain features, frequency domain features, and time-frequency features, and use correlation analysis and feature importance ranking algorithms to screen out the most representative feature subset for fault identification; S3. Input the feature subset into a pre-trained hybrid fault detection model. Detect abnormal behavior and generate preliminary fault alerts based on the unsupervised learning model. Verify and classify the alerts using the supervised learning model to determine the fault type. S4. Use a hierarchical diagnostic strategy to locate the subsystem where the fault is located. Build a PV system topology model using a graph neural network. Combined with a Bayesian inference algorithm, analyze the fault propagation path, determine the root cause of the fault, and generate a diagnostic report that includes the fault location, type, severity, and repair recommendations. S5. Based on newly emerged fault cases and system operation data, regularly update the fault feature library and detection model, and use transfer learning and incremental learning techniques to improve the generalization ability of the model.

2. A photovoltaic system fault identification method according to claim 1, characterized in that: The specific screening method of the feature subset is as follows: Calculate the Pearson correlation coefficient matrix between time domain features, frequency domain features, and time-frequency features, identify feature pairs whose absolute value of the correlation coefficient exceeds a preset threshold, and calculate the correlation between each pair of highly correlated features and the fault label. Retain features with higher correlation to the fault label and remove redundant features. Use the random forest algorithm to score the importance of the features screened by correlation analysis, and sort the features in descending order based on the feature importance scores; A preset number of top-level important features are selected from the sorted features to form the final feature subset.

3. A photovoltaic system fault identification method according to claim 2, characterized in that: The unsupervised learning algorithm adopts the isolation forest algorithm, and the supervised learning algorithm adopts the random forest classifier.

4. A photovoltaic system fault identification method according to claim 3, characterized in that: In step S3, the specific method for determining the fault type is as follows: Input the feature subset into the isolation forest algorithm, calculate the anomaly score of each sample, dynamically set the anomaly threshold based on the statistical characteristics of historical data, and generate a preliminary fault alert when the sample anomaly score exceeds the threshold; Input the sample features that trigger the alarm into a random forest classifier, which is trained with multiple types of fault samples and outputs a probability distribution of fault types; If the highest probability category score exceeds the preset confidence threshold, it is confirmed as the corresponding fault type; If the threshold is not exceeded, the fuzzy decision-making mechanism is activated to make a secondary judgment based on the anomaly score, environmental parameters, and equipment operating status; Based on real-time environmental parameters and system operating status, the anomaly threshold of the isolation forest is dynamically corrected through Gaussian process regression to reduce the false alarm rate caused by environmental fluctuations.

5. A photovoltaic system fault identification method according to claim 4, characterized in that: In step S4, the specific method for determining the root cause of the fault is as follows: Based on the fault type output by the hybrid fault detection model, the preset fault-subsystem mapping matrix is ​​matched to preliminarily determine the set of candidate subsystems that may contain faults; For the PV strings within the candidate subsystem, a spatiotemporal correlation network of electrical parameters is constructed to calculate the similarity of operating states between strings and identify abnormal string clusters that deviate from normal patterns. For abnormal string clusters, device-level time series features are extracted and input into a trained graph neural network model. The graph neural network model builds node-edge relationships based on the PV system topology and infers the fault propagation path through a message passing mechanism. By integrating the failure probability distribution output by the graph neural network model, historical equipment failure rate data, and real-time environmental factors, Bayesian reasoning is used to calculate the posterior failure probability of each device. For devices whose posterior probability exceeds the confidence threshold, the adjacent device status consistency check is initiated. If a cascading anomaly exists, it is identified as the fault source. Otherwise, the diagnostic path is retroactively adjusted. Generate a topological heat map that includes fault propagation paths, impact ranges, and root cause probabilities, and mark key device nodes recommended for inspection.

6. A photovoltaic system fault identification method according to claim 1, characterized in that: The environmental data includes light intensity, temperature, humidity and wind speed, the electrical parameters include voltage, current and power, and the device status data includes inverter working status and component temperature difference.

7. The photovoltaic system fault identification method according to claim 1, characterized in that: The time domain features include mean, variance and peak value, the frequency domain features include power spectrum density and frequency components, and the time-frequency features include wavelet transform coefficients.

8. A photovoltaic system fault identification system, used to implement the photovoltaic system fault identification method according to any one of claims 1 to 7, characterized in that: include: The data acquisition module includes environmental sensors, electrical parameter sensors, and equipment status monitoring sensors, which are used to collect environmental data, electrical parameter data, and equipment status data in real time and transmit the data to the data processing module; The data processing module is used to clean and standardize the collected raw data, extract time domain features, frequency domain features and time-frequency features, and perform feature screening to obtain feature subsets; Fault detection and classification module, which includes pre-trained isolation forest algorithm and random forest classifier, is used to receive feature subsets and perform anomaly detection, preliminary fault alert generation, fault type verification and classification; The fault location and root cause analysis module uses graph neural networks and Bayesian reasoning algorithms to perform fault location and root cause analysis based on the fault type and generate a diagnostic report. The model update module regularly updates the fault feature library and detection model based on new fault cases and system operation data, and uses transfer learning and incremental learning technologies to optimize model performance.

Images