Photovoltaic operation fault diagnosis method and system

Abstract

The invention provides a photovoltaic operation fault diagnosis method and system, and the method comprises the steps: constructing a fault mode and influence factor database based on real-time monitoring data and historical record data, carrying out the feature extraction, obtaining a fault feature vector, and constructing a fault diagnosis model through a preset decision algorithm; performing label classification on the fault feature vector to obtain a label classification data set, and constructing a label classification model; constructing a fault propagation model based on the fault mode and the influence factor database; and based on the plurality of models, performing fault diagnosis analysis on the real-time monitoring data to obtain a fault diagnosis report. According to the method, the problem that the operation reliability of a photovoltaic system is poor due to the fact that only processing is carried out when abnormity occurs in the whole operation state and performance expression and further analysis and processing are not carried out on the propagation path and potential influence of the abnormity or the fault in the prior art is solved. The fault condition of the system is diagnosed and analyzed by constructing a plurality of models, and stable operation of the system is ensured.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and in particular to a photovoltaic operation fault diagnosis method and system. Background Art

[0002] When a photovoltaic system fails, it is usually in a state of multiple faults coexisting, such as photovoltaic panel damage, inverter failure, and cable connection problems. These faults may not only exist independently, but may also affect and couple with each other to form a complex fault combination. For complex fault combinations, if the faults cannot be effectively analyzed and processed, other components will fail one after another, causing the propagation and diffusion of the faults, expanding the scope of the fault impact and seriously affecting the stable operation of the system.

[0003] At present, the existing technology has proposed a low-voltage distributed photovoltaic panoramic monitoring method to address the existing problems. After processing and analyzing the fault data information, it predicts the possible problems of low-voltage distributed photovoltaic power generation and reminds relevant personnel to perform preventive maintenance. However, the existing problem is that the existing technology only processes the abnormalities in the overall operating status and performance of the low-voltage distributed photovoltaics, without further analyzing and processing the propagation path of the abnormality or fault and the potential impact, and there is still the problem of poor reliability of photovoltaic system operation. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a photovoltaic operation fault diagnosis method and system, which fully considers the propagation path of the abnormality or fault and further analyzes and processes the potential impact.

[0005] To achieve the above-mentioned purpose, an embodiment of the present invention provides a photovoltaic operation fault diagnosis method, comprising: constructing a fault mode and influencing factor database based on real-time monitoring data and historical record data of the photovoltaic system;

[0006] Based on a preset feature extraction algorithm, feature extraction is performed on the fault mode and influencing factor database to obtain a fault feature vector;

[0007] Based on the fault feature vector, a fault diagnosis model is constructed through a preset decision algorithm;

[0008] Based on a preset classification algorithm, label classification is performed on the fault feature vector to obtain a label classification data set;

[0009] Based on the label classification data set, construct a label classification model;

[0010] Based on the fault mode and influencing factor database, a fault propagation model is constructed by simulating the system operation state and fault propagation path through a preset neural model;

[0011] Based on the fault diagnosis model, label classification model and fault propagation model, fault diagnosis and analysis are performed on the real-time monitoring data to obtain a photovoltaic operation fault diagnosis report.

[0012] The embodiment of the present invention proposes a photovoltaic operation fault diagnosis method, which ensures the reliability of the data source by acquiring real-time monitoring data and historical record data, and provides a data basis for subsequent fault diagnosis; further improves the accuracy of the data by extracting features from the data and constructing feature vectors; obtains the association between different fault modes and identifies the key features of each fault mode by constructing a fault diagnosis model, so as to improve the accuracy of fault identification and diagnosis; then, classifies the fault feature vectors by labels and constructs a label classification model, identifies and distinguishes various fault conditions in the system, so as to improve the accuracy of fault analysis; further, constructs a fault propagation model by simulating the system operation state and the fault propagation path, obtains data on the propagation of the fault from the source, analyzes the fault propagation path and the potential impact of the fault on other components, so as to improve the reliability of fault diagnosis; finally, diagnoses and analyzes the fault conditions of the system by constructing multiple models, generates an effective fault diagnosis report, and assists in taking preventive measures in advance to block the propagation of faults and ensure the stable operation of the system.

[0013] Furthermore, the construction of a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system includes:

[0014] Based on the preset deep learning algorithm, semantic extraction and several semantic features are performed on real-time monitoring data and historical record data;

[0015] Based on a preset clustering algorithm, cluster analysis is performed on the plurality of semantic features to obtain different types of failure modes;

[0016] Based on a preset association rule mining algorithm, association factor mining is performed on the historical record data and different types of failure modes to obtain an influencing factor database.

[0017] Furthermore, the feature extraction of the fault mode and influencing factor database based on a preset feature extraction algorithm to obtain a fault feature vector includes:

[0018] Based on a preset feature extraction algorithm, statistical features, trend features and periodic features of the failure mode and influencing factor database are extracted respectively to construct a feature group;

[0019] The feature groups are combined to obtain a fault feature vector.

[0020] Furthermore, the fault diagnosis model is constructed based on the fault feature vector by a preset decision algorithm, including:

[0021] Based on preset support and confidence thresholds, performing correlation analysis on the fault feature vector to obtain a correlation analysis result;

[0022] Based on the correlation analysis result, the correlation between each feature vector and the fault mode is evaluated, and a key feature subset that meets the preset correlation requirement is selected;

[0023] Based on a preset decision algorithm, a fault diagnosis model is constructed with the key feature subset as input and the relevant fault types as output.

[0024] Furthermore, based on a preset classification algorithm, the fault feature vector is subjected to label classification to obtain a label classification data set, including:

[0025] Based on the preset classification algorithm, the fault feature vector is labeled with multiple labels to obtain several multi-label fault feature vectors;

[0026] By classifying the fault types of the plurality of multi-label fault feature vectors, a label classification data set is constructed.

[0027] Furthermore, based on the label classification data set, a label classification model is constructed, including:

[0028] An initial label classification model is constructed with the feature vector as input and the fault type as output;

[0029] The initial label classification model is trained using the label classification data set to obtain a model training evaluation index;

[0030] Based on the model training evaluation index, the model parameters and model structure of the initial label classification model are optimized until the preset performance requirements are met to obtain a label classification model.

[0031] Furthermore, based on the fault mode and influencing factor database, a fault propagation model is constructed by simulating the system operation state and fault propagation path through a preset neural model, including:

[0032] Inputting the fault mode and influencing factor database into a preset neural model for training to obtain simulation data of fault types and fault factors;

[0033] Based on the simulation data of the fault type and the fault factor, simulate the system operation state and the fault propagation path to obtain the fault propagation simulation data;

[0034] A fault propagation model is constructed based on the fault propagation simulation data.

[0035] Furthermore, based on the fault diagnosis model, label classification model and fault propagation model, the real-time monitoring data is analyzed for fault diagnosis to obtain a photovoltaic operation fault diagnosis report, including:

[0036] Based on the fault diagnosis model, the real-time monitoring data is subjected to fault trigger judgment to obtain a fault judgment result;

[0037] Based on the label classification model, label classification is performed on the fault judgment result to obtain a fault classification result;

[0038] Based on the fault propagation model, simulating the fault propagation path of the fault classification result to obtain a fault propagation result;

[0039] Based on the fault judgment results, fault classification results and fault propagation results, a fault diagnosis report is output.

[0040] Furthermore, the method further comprises the following steps:

[0041] Based on a preset knowledge graph framework, a number of fault diagnosis data are obtained by reading data from the fault diagnosis report;

[0042] Performing weight configuration on the plurality of fault diagnosis data to obtain fault diagnosis feedback data;

[0043] The fault diagnosis feedback data is input into the fault diagnosis model, the label classification model and the fault propagation model for feedback training to update the fault diagnosis model, the label classification model and the fault propagation model.

[0044] A photovoltaic operation fault diagnosis method proposed in an embodiment of the present invention integrates fault diagnosis data into a knowledge graph, a fault diagnosis model, the label classification model and the fault propagation model for feedback training, and continuously updates the model to adapt to system evolution and newly discovered fault modes, including updating high-weight content in fault coexistence and fault chains, and then retrains the model to adapt to system evolution and new fault modes to ensure system operation reliability.

[0045] The embodiment of the present invention also provides a photovoltaic operation fault diagnosis system, comprising: a data construction module, a feature extraction module, a first model construction module, a data classification module, a second model construction module, a third model construction module and a fault diagnosis module;

[0046] The data construction module is used to construct a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system;

[0047] The feature extraction module is used to extract features from the fault mode and influencing factor database based on a preset feature extraction algorithm to obtain a fault feature vector;

[0048] The first model building module is used to build a fault diagnosis model based on the fault feature vector through a preset decision algorithm;

[0049] The data classification module is used to perform label classification on the fault feature vector based on a preset classification algorithm to obtain a label classification data set;

[0050] The second model building module is used to build a label classification model based on the label classification data set;

[0051] The third model building module is used to build a fault propagation model by simulating the system operation state and fault propagation path through a preset neural model based on the fault mode and influencing factor database;

[0052] The fault diagnosis module is used to perform fault diagnosis analysis on real-time monitoring data based on a fault diagnosis model, a label classification model and a fault propagation model to obtain a photovoltaic operation fault diagnosis report.

[0053] The embodiment of the present invention proposes a photovoltaic operation fault diagnosis system, which obtains real-time monitoring data and historical record data through a data construction module to ensure the reliability of the data source and provide a data basis for subsequent fault diagnosis; extracts features from the data through a feature extraction module to construct a feature vector to further improve the accuracy of the data; constructs a fault diagnosis model through a first model construction module to obtain the association between different fault modes and identify the key features of each fault mode, so as to improve the accuracy of fault identification and diagnosis; then, the fault feature vectors are respectively labeled and classified by a data classification module and a second model construction module to construct a label classification model, identify and distinguish various fault conditions in the system, so as to improve the accuracy of fault analysis; further, the system operation state and the fault propagation path are simulated through a third model construction module, a fault propagation model is constructed, data of the fault propagation from the source is obtained, the fault propagation path and the potential impact of the fault on other components are analyzed, so as to improve the reliability of fault diagnosis; finally, the fault diagnosis module uses multiple models to diagnose and analyze the fault conditions of the system, generates an effective fault diagnosis report, and assists in taking preventive measures in advance to block the propagation of faults and ensure the stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic flow chart of the steps of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention;

[0055] Figure 2 A module structure diagram of a photovoltaic operation fault diagnosis system provided by an embodiment of the present invention;

[0056] Figure 3 A schematic flow chart of feedback training steps of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention;

[0057] Figure 4 A schematic flow chart of label classification steps of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention;

[0058] Figure 5 A schematic diagram of a troubleshooting process of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention;

[0059] Figure 6 A schematic flow chart of fault diagnosis steps combined with model updating in a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Example 1

[0062] See also Figure 1 , Figure 1 A schematic flow chart of the steps of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention. Figure 1 As shown, the embodiment of the present invention provides a photovoltaic operation fault diagnosis method, including steps 101 to 107, each step is specifically as follows:

[0063] Step 101, constructing a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system;

[0064] As an example of this embodiment, based on a preset deep learning algorithm, semantic extraction is performed on real-time monitoring data and historical record data to obtain a number of semantic features; based on a preset clustering algorithm, cluster analysis is performed on the several semantic features to obtain different types of failure modes; based on a preset association rule mining algorithm, association factor mining is performed on the historical record data and different types of failure modes to obtain an influencing factor database.

[0065] A specific implementation method is to obtain the type, material composition, life information, real-time monitoring data such as voltage, current, temperature, power output, fault history record and maintenance log of photovoltaic cells, construct a multi-dimensional fault mode feature vector, and use a deep learning algorithm to extract features and represent the multi-dimensional fault mode feature vector to obtain a low-dimensional semantic representation; then use a clustering algorithm to perform clustering analysis on the low-dimensional semantic representation to identify different types of failure modes; finally, through association rule mining algorithms such as Apriori, FP-Growth and Eclat, combined with fault history records and maintenance logs, obtain the correlation and influencing factors between different failure modes and build an influencing factor database. For a more specific example, by extracting features and learning representations of fault data from 1,000 photovoltaic panels, the original 500-dimensional fault feature vector was reduced to a 50-dimensional semantic representation vector; the K-means clustering algorithm was used to cluster the 50-dimensional semantic representations of 10,000 fault modes, and five main types of fault modes were identified; by mining association rules of 20,000 fault history records and maintenance logs, appropriate support and confidence thresholds were set, and strong association rules that met the threshold conditions were found. Finally, association results were obtained, such as the association support of the "battery cell damage" fault mode and the "connector damage" fault mode was 2%, and the confidence was 60%. Through several association results, a database of influencing factors was constructed.

[0066] Step 102, based on a preset feature extraction algorithm, feature extraction is performed on the fault mode and influencing factor database to obtain a fault feature vector;

[0067] As an example of this embodiment, based on a preset feature extraction algorithm, statistical features, trend features and periodic features of the fault mode and influencing factor database are extracted respectively to construct feature groups; the feature groups are combined to obtain a fault feature vector.

[0068] A specific implementation method is to obtain real-time monitoring data of the photovoltaic system, including: temperature, power output, inverter current and voltage, etc. of the photovoltaic panel, first clean the obtained data, remove missing values ​​and outliers, and use methods such as minimum-maximum normalization or Z-score normalization to scale the data to a preset range so that the data meets the analysis requirements; a specific preferred solution is that in the data cleaning process, methods such as mode and median can be used to fill in the missing values ​​of monitoring indicators such as temperature and voltage, and the Laida criterion (3σ principle) can be used to identify and remove outliers that exceed the normal range. The outliers in the range are then normalized using the minimum-maximum normalization method to scale the temperature data to the range of 0-100 degrees Celsius and the voltage data to the range of 0-1000 volts. In feature engineering, statistical indicators such as the mean, standard deviation, and 90% quantile of the temperature data can be calculated. The EWMA (exponentially weighted moving average) model is used to extract the long-term trend characteristics of the temperature, and the discrete Fourier transform is used to extract the periodic characteristics of the temperature (such as a 24-hour cycle). For non-time series data such as current and power, derived features such as peak current and average power can be constructed based on physical properties.

[0069] The cleaned data is processed as follows:

[0070] (1) Use feature engineering methods to extract statistical characteristics of indicators such as temperature, power, current, and voltage, such as mean, variance, peak, percentile, kurtosis, and skewness;

[0071] (2) For time series data, use methods such as moving average and exponential smoothing to extract trend characteristics;

[0072] (3) Extract periodic features using methods such as FFT and wavelet transform;

[0073] The above extracted features are combined into a high-dimensional feature vector, namely, a fault feature vector;

[0074] Step 103, based on the fault feature vector, a fault diagnosis model is constructed by a preset decision algorithm;

[0075] As an example of this embodiment, based on preset support and confidence thresholds, the fault feature vector is subjected to correlation analysis to obtain a correlation analysis result; based on the correlation analysis result, the correlation between each feature vector and the fault mode is evaluated, and a key feature subset that meets the preset correlation requirements is screened; based on a preset decision algorithm, a fault diagnosis model is constructed with the key feature subset as input and relevant fault types as output.

[0076] A specific implementation method is to set appropriate support and confidence thresholds, analyze the correlation of fault feature vectors, obtain the correlation between temperature anomaly and power drop, the correlation between current fluctuation and voltage fluctuation, etc., and then use feature selection methods such as mutual information and relief to evaluate the correlation between each feature and the fault mode. The importance of features is measured using indicators such as information gain ratio and Gini index. Through methods such as recursive feature elimination, the key feature subset with the greatest contribution to fault diagnosis is gradually screened out. The screened key features (equivalent to the key feature subset) are used as input, and the fault types of historical fault data (equivalent to related fault types) are labeled as output. A decision tree is used to construct an A I-driven fault diagnosis model; the preferred solution is to use the Apriori algorithm for association rule mining, set the minimum support to 0.05, the minimum confidence to 0.8, and the maximum item set length to 3, find strong association rules such as "abnormal high temperature & low power output -> photovoltaic panel failure", use the Relief-F algorithm for feature selection, set the nearest neighbor number k=10, the weight threshold to 0.01, select the top 20 key features such as temperature mean and power peak, and use the CART (classification and regression tree) algorithm to build an AI-driven fault diagnosis model. The Gini coefficient is used as the metric for node splitting, the maximum depth of the tree is set to 5, and the minimum number of leaf node samples is set to 10.

[0077] As another example of this embodiment, by recursively dividing the feature space, a decision rule for fault diagnosis is constructed. During the growth of the decision tree, the Gini coefficient and information gain ratio are used to select the optimal division attribute. The complexity of the tree is controlled by setting parameters such as the maximum depth and the minimum number of samples to avoid overfitting. The k-fold cross-validation method is used to divide the data set into k parts, which are used as validation sets and training sets in turn to optimize the generalization performance of the model. Pre-pruning or post-pruning techniques are used to cut off some branches according to evaluation indicators during the tree growth process or after the complete decision tree is generated to further simplify the model. The trained AI-driven fault diagnosis model is tested and evaluated, and multiple indicators such as accuracy, precision, recall rate, F1-score, ROC curve and AUC value are used to comprehensively evaluate the fault diagnosis performance of the model. To further improve the performance, you can try ensemble learning methods such as random forest and AdaBoost, combine multiple decision tree models for prediction, until the model performance meets the preset requirements, and obtain the fault diagnosis model with the best performance. The preferred solution is to use 5-fold cross-validation, with an average accuracy of more than 85%. Use the cost-complexitypruning algorithm for post-pruning, adjust the penalty coefficient α on the validation set, and simplify the model complexity to the greatest extent while maintaining the performance of the decision tree. The final model has a fault diagnosis accuracy of 90% on the test set, and both the precision and recall rates exceed 85%, with an AUC value of more than 0.95. In order to further improve the performance, try using the random forest algorithm and integrate 50 decision trees, and the average accuracy can be increased by 2-3 percentage points.

[0078] Step 104, based on a preset classification algorithm, label classification is performed on the fault feature vector to obtain a label classification data set;

[0079] As an example of this embodiment, see Figure 4 , Figure 4 A schematic flow chart of a label classification step of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown, based on a preset classification algorithm, the fault feature vector is annotated with multiple labels to obtain a number of multi-label fault feature vectors; and a label classification data set is constructed by classifying the fault types of the multi-label fault feature vectors.

[0080] A specific implementation method is to label different types of faults such as photovoltaic panel damage and inverter faults based on the historical operation data of the photovoltaic system and the description in the fault record, and construct a multi-label data set, in which each sample may belong to multiple fault categories at the same time. If a sample has both photovoltaic panel damage and inverter fault, it is labeled as [1, 1], otherwise it is labeled as [1, 0] or [0, 1], thereby obtaining several multi-label fault feature vectors. Then, by classifying the fault types of the several multi-label fault feature vectors, a label classification data set is constructed.

[0081] Step 105, constructing a label classification model based on the label classification data set;

[0082] As an example of this embodiment, see Figure 4 , Figure 4 A schematic flow chart of a label classification step of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown, an initial label classification model is constructed with a feature vector as input and a fault type as output; the initial label classification model is trained using the label classification data set to obtain a model training evaluation index; based on the model training evaluation index, the model parameters and model structure of the initial label classification model are optimized until the preset performance requirements are met, thereby obtaining a label classification model.

[0083] A specific implementation method is to use deep learning models such as multi-layer perceptron (MLP) and convolutional neural network (CNN), take the extracted feature vector as input and the fault type as output to build a multi-label classification model; when designing the model, use Sigmoid activation function instead of Softmax, regard each category as an independent binary classification problem, and predict the probability that the sample belongs to each fault category. In the model training process, use a multi-label version of the binary cross entropy loss function, that is, regard each category as an independent binary classification problem, calculate the cross entropy loss of all categories for each sample, and then take the average as the overall loss. Use Adam optimization algorithm to update model parameters. In order to alleviate the category imbalance To solve the problem of balancing the classification of faults, SMOTE (synthetic minority oversampling technology) is used to oversample the minority class samples and generate new synthetic samples. Random undersampling can also be used to randomly select a subset of the majority class samples with the same number of minority class samples to balance the category distribution of the data set. At the same time, a weight coefficient is introduced into the loss function during training to give a higher weight to the fault category with fewer samples. When evaluating the model, multi-label classification-specific evaluation indicators such as HammingLoss and ExactMatchRatio are used to comprehensively measure the performance of the model in each category. Through hyperparameter optimization methods such as grid search and random search, the model structure and training parameters are adjusted to obtain the model with the best performance.

[0084] As another example of this embodiment, an isolation forest algorithm is used to detect anomalies in current and temperature data, and the contamination parameter is set to 0.1, that is, there are 10% abnormal samples in the data set. Through anomaly detection, the characteristic pattern when the photovoltaic panel damage and the inverter fault coexist is identified, such as the current fluctuation range exceeds 2 standard deviations and the temperature rise exceeds 10 degrees Celsius. Specifically, for the situation where the photovoltaic panel damage and the inverter fault coexist, the abnormal pattern of the current and temperature data is focused on, and a multivariate anomaly detection algorithm is used to comprehensively consider the correlation of multiple indicators to identify the characteristic pattern when these two faults coexist, such as multivariate Gaussian distribution and matrix decomposition. Deep learning models such as multilayer perceptron (MLP) and convolutional neural network (CNN) are used, and the extracted feature vector is used as input and the fault type is used as output to train a multi-label classification model. When training the MLP model, a network structure including 3 hidden layers can be designed, and each layer contains 128, 64 and 32 neurons respectively. ReL U activation function and Dropout regularization, the Dropout rate is set to 0.5; when training the CNN model, 2 convolutional layers can be used, the first layer uses 16 5x5 convolution kernels, the second layer uses 32 3x3 convolution kernels, the convolution layer is followed by the MaxPooling layer and the BatchNormalization layer, and then 2 fully connected layers are used for classification; when using SMOTE for oversampling, the number of minority class samples can be increased to 50% of the majority class samples, and then merged with the original data set to form a more balanced data set; when evaluating the model, 5-fold cross validation is used to calculate the HammingLoss and F1-score of the model on each fold, and then the average is taken as the final performance indicator; the hyperparameters of the MLP model such as the learning rate, the number of hidden layers, and the batch size are adjusted through grid search to obtain the model with the lowest HammingLoss on the validation set, and the F1-score of this model on the test set can reach above 0.85.

[0085] Step 106, based on the fault mode and influencing factor database, simulate the system operation status and fault propagation path through a preset neural model to construct a fault propagation model; as an example of this embodiment, input the fault mode and influencing factor database into the preset neural model for training to obtain simulation data of fault types and fault factors; based on the simulation data of the fault types and fault factors, simulate the system operation status and fault propagation path to obtain fault propagation simulation data; based on the fault propagation simulation data, construct a fault propagation model.

[0086] A specific implementation method is to collect real-time operating data of each component of the photovoltaic system, including photovoltaic panel temperature, inverter voltage and current, and photovoltaic module power output, as input parameters of the dynamic simulation module, and simulate the system operating status and fault propagation path of photovoltaic panel temperature, inverter voltage and current, and photovoltaic module power output, as follows:

[0087] (1) Using the photovoltaic panel temperature-performance module, the performance changes of photovoltaic panels under different temperature conditions are simulated to obtain the data on the impact of temperature on the performance of photovoltaic panels;

[0088] (2) Using the inverter voltage and current response module to simulate the inverter response under voltage and current fluctuations, and obtain data on the impact of voltage and current fluctuations on inverter operation;

[0089] (3) Simulate the impact of dust and shading on the power output of photovoltaic modules and obtain the power drop data of photovoltaic modules caused by dust shading.

[0090] By integrating the temperature performance change data of photovoltaic panels, the voltage and current response data of inverters, and the power drop data of photovoltaic components due to dust shading, a photovoltaic system fault propagation model is constructed through a deep learning neural network.

[0091] As another example of this embodiment, real-time operating data of each component of the photovoltaic system is obtained, such as the photovoltaic panel temperature of 45°C, the inverter voltage of 220V and the current of 5A, the photovoltaic module power output of 1000W, etc., as input parameters of dynamic simulation, according to the photovoltaic panel temperature data, the photovoltaic panel temperature-performance module is used to simulate the performance changes of the photovoltaic panel under different temperature conditions, such as when the temperature rises by 10°C, the output power of the photovoltaic panel decreases by 5%, and the temperature The impact of photovoltaic panel performance data is obtained; according to the inverter voltage and current data, the inverter voltage and current-response module is used to simulate the response of the inverter under voltage and current fluctuations. For example, if the voltage fluctuation exceeds 10% of the rated value, the inverter protection device is activated and stops working, and the voltage and current fluctuations. The impact of the inverter operation is obtained; according to the photovoltaic module power output data, the impact of dust and shielding on the photovoltaic module power output is simulated. For example, for every 10 mg / m2 increase in dust density, the output power of the photovoltaic module decreases by 1%, and the power drop data of the photovoltaic module caused by dust shielding is obtained; by integrating the temperature performance change data of the photovoltaic panel, the voltage and current response data of the inverter, and the power drop data of the photovoltaic module due to dust shielding, a photovoltaic system fault propagation model is constructed. For example, when the temperature of the photovoltaic panel exceeds 60°C, the output power drops by 20%, which in turn causes the inverter voltage and current fluctuations, and finally causes the inverter protection device to operate, simulating the propagation process of the fault from the source; the photovoltaic system fault propagation model is used to analyze the propagation paths of faults such as abnormal photovoltaic panel temperature, inverter voltage and current fluctuations, and dust shielding of photovoltaic modules. For example, the probability of abnormal photovoltaic panel temperature causing inverter voltage and current fluctuations is 30%, and the probability of inverter voltage and current fluctuations causing the protection device to operate is 20%, and the potential impact of each fault on other components of the photovoltaic system is determined.

[0092] Step 107 , based on the fault diagnosis model, the label classification model and the fault propagation model, a fault diagnosis analysis is performed on the real-time monitoring data to obtain a photovoltaic operation fault diagnosis report.

[0093] As an example of this embodiment, based on the fault diagnosis model, a fault trigger judgment is performed on the real-time monitoring data to obtain a fault judgment result; based on the label classification model, the fault judgment result is labeled and classified to obtain a fault classification result; based on the fault propagation model, a fault propagation path simulation is performed on the fault classification result to obtain a fault propagation result; based on the fault judgment result, the fault classification result and the fault propagation result, a fault diagnosis report is output.

[0094] For a specific implementation method, see Figure 5 , Figure 5 A schematic diagram of a troubleshooting process of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 5As shown in the figure, the real-time monitoring data of the photovoltaic system is analyzed through the AI-driven fault diagnosis model to determine whether there is a fault. If a fault is detected, the fault diagnosis process is triggered to make a preliminary judgment on the fault type, such as broken photovoltaic panel connection lines, water ingress into the junction box, blown series fuses, poor inverter heat dissipation, aging of DC side capacitors, and burnout of AC side relays. Relevant data within the time period when the fault occurred is extracted from the monitoring database, including the temperature, irradiance and power output of the photovoltaic panel, the voltage, current and efficiency of the inverter, and environmental factors such as wind speed and humidity, to form a structured data set. The data set contains fields such as timestamp, device ID, fault type and measurement value, which is convenient for subsequent analysis and modeling. The fault data is input into the transparent interpretation component, and the contribution of each influencing factor to the fault is calculated through feature importance analysis algorithms such as PermutationImportance and SHAP. Among them, PermutationImportance randomly disrupts the feature values ​​and evaluates the degree of model performance degradation. Measure feature importance; SHAP explains model behavior by calculating the contribution of each feature to the model output. At the same time, it uses decision rule extraction algorithms, such as DecisionTree and AssociationRule, to extract if-then rules from the trained fault diagnosis model, revealing the association between features and fault types, and obtaining quantitative explanation results. For example, abnormal temperature (over 80°C) causes a 30% decrease in photovoltaic panel efficiency, and excessive wind speed (over 20m / s) causes an inverter overload of 20%. Based on the fault diagnosis and explanation results, a fault diagnosis report is generated, listing the specific types of photovoltaic panel damage and inverter failure, and providing cause analysis and solution suggestions.The report also includes some quantitative indicators, such as a drop in photovoltaic panel efficiency of more than 20%, an inverter conversion efficiency of less than 90%, etc., to more accurately describe the severity of the fault. At the same time, natural language generation technology is used to convert structured diagnostic data into easy-to-understand text descriptions, and the diagnostic results and influencing factors are presented in the report in a visual way such as charts and graphs. For example, a bar chart is used to show the proportion of each fault type, a line chart is used to show the trend of temperature and power changes over time, and a radar chart is used to show the contribution of each influencing factor. The characteristics and laws of the fault are intuitively revealed to obtain the fault judgment results; the trained multi-label classification model is deployed to the production environment, and real-time fault diagnosis is performed on the newly collected photovoltaic system operation data. When the model predicts that the sample may belong to multiple fault categories, it is sorted according to the predicted probability of each category, and the fault type with a predicted probability greater than the preset threshold is selected as For the final diagnosis result, if it is necessary to control the number of fault types in the report, a threshold can be set to select only the types with a predicted probability greater than the threshold to obtain the fault classification result; the photovoltaic system fault propagation model is used to analyze the propagation paths of faults such as abnormal photovoltaic panel temperature, inverter voltage and current fluctuations, and dust shielding of photovoltaic components, and determine the potential impact of each fault on other components of the photovoltaic system. According to the fault propagation path and impact analysis results, the association rule mining algorithm is used to find the correlation between different faults and obtain the mutual influence law between faults. According to the mutual influence law of faults, a photovoltaic system fault prediction model is constructed through a vector machine algorithm. By real-time monitoring of photovoltaic system operation data, possible chain failures are predicted, the trend of fault propagation is judged, and the fault propagation results are obtained; according to the above-mentioned fault judgment results, fault classification results and fault propagation results, a fault diagnosis report is output.

[0095] As another example of this embodiment, by performing feature engineering on historical monitoring data, 20 indicators such as photovoltaic panel temperature, power and efficiency are extracted as fault diagnosis model inputs, and fault types such as photovoltaic panel open circuit and string short circuit are used as model outputs. The hyperparameters of the decision tree, such as the maximum depth and the minimum number of leaf node samples, are optimized using the grid search method, and the model performance is evaluated through 5-fold cross validation, and the fault diagnosis accuracy can reach more than 95%. When new monitoring data arrives, it is passed into the trained fault diagnosis model, and the fault type determination result can be obtained within 10 milliseconds. For structured monitoring data sets, the Pandas library can be used for data cleaning and processing. Through outlier detection algorithms such as Z-score and LOF, outliers of indicators such as temperature and voltage can be identified and filtered out to improve data quality. Data interpolation algorithms such as linear interpolation and spline interpolation are used to fill in missing data. Time series feature extraction algorithms such as moving average and first-order difference are used to mine features in the time dimension, such as temperature rise rate and power fluctuation period. When using the SHAP algorithm to explain the model, the shap.TreeExplainer interface can be called to pass in the model and test set features to calculate the SHAP value of each feature. The SHAP values ​​are plotted as ForcePlot to intuitively show the positive and negative contributions of each feature to the fault. According to the sorting of SHAP values, the five features with the highest contribution are screened out as explanations of the cause of the fault. Natural language processing libraries such as nltk and spaCy are used to perform semantic analysis and keyword extraction on the fault cause and handling suggestions, and then fill them into predefined templates to generate easy-to-understand fault descriptions and handling processes. The whole process takes no more than 1 second.Visualization of fault reports can be achieved using drawing libraries such as Matplotlib and Seaborn. Through the subplot function of Matplotlib, multiple subgraphs can be drawn in the same graph, such as a bar graph showing the distribution of fault types and a line graph showing the trend of fault time. Appropriate legends, titles, and axis labels can be set to beautify the chart style. The heatmap function of Seaborn can be used to draw a correlation heat map of the fault causes and analyze the interaction between the various influencing factors. The mpld3 library is used to convert static charts into interactive web pages, which supports operations such as zooming, hovering, and animation to enhance the information transmission effect of the chart. Finally, the FlaskWeb framework is used to encapsulate the fault analysis process into a RESTfulAPI interface and integrate it with the company's existing operation and maintenance platform. The fault diagnosis model is deployed on the Flask server. Through the HTTPPOST method, monitoring data in JSON format is received to predict the fault type online. The diagnostic report is sent to the operation and maintenance personnel's mailbox through the SMTP protocol and pushed to the operation and maintenance screen in real time through WebSocket. The RabbitMQ message queue is used to ensure the reliable transmission of fault information and avoid data loss caused by network jitter. dis cache, stores hot spot fault data, improves the efficiency of data query and display, the throughput of the whole system can reach 100 diagnostic requests per second, and the average response delay is controlled within 1 second; according to the fault propagation path and impact analysis results, the association rule mining algorithm is adopted, such as the correlation between the abnormal temperature of the photovoltaic panel and the voltage and current fluctuation of the inverter is 8, and the correlation between the voltage and current fluctuation of the inverter and the action of the protection device is 6, the correlation between different faults is found, and the mutual influence law between the faults is obtained; the label classification model is used to predict the newly collected photovoltaic system operation data. If the predicted probability of the fault type of a sample exceeds 0.7, It is added to the fault diagnosis report to prompt the operation and maintenance personnel to conduct investigation and repair. At the same time, the association rules of fault types can also be analyzed. For example, the confidence level of "inverter fault->photovoltaic panel damage" is as high as 0.8, indicating that inverter failure is an important cause of photovoltaic panel damage. According to the law of mutual influence of faults, a photovoltaic system fault prediction model (equivalent to a fault propagation model) is constructed. By real-time monitoring of photovoltaic system operation data, such as the current photovoltaic panel temperature is 55°C, the probability of inverter voltage and current fluctuation is predicted to be 60%, and the probability of triggering the action of the protection device is 30%. Possible chain failures are predicted and the trend of fault propagation is judged.

[0096] As another example of this embodiment, see Figure 3 , Figure 3 A schematic flow chart of a feedback training step of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 3As shown, a photovoltaic operation fault diagnosis method proposed in an embodiment of the present invention further includes the following steps, specifically:

[0097] Step 301, based on a preset knowledge graph framework, obtain a number of fault diagnosis data by reading data from the fault diagnosis report;

[0098] Step 302, weighting the plurality of fault diagnosis data to obtain fault diagnosis feedback data;

[0099] Step 303: input the fault diagnosis feedback data into the fault diagnosis model, the label classification model and the fault propagation model for feedback training to update the fault diagnosis model, the label classification model and the fault propagation model.

[0100] See also Figure 6 , Figure 6 A schematic flow chart of a fault diagnosis step combined with model updating of a photovoltaic operation fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 6As shown, according to the actual fault handling and diagnosis results, the feedback information is integrated into the knowledge graph and fault diagnosis model, and the model is continuously updated to adapt to the system evolution and newly discovered fault modes, including updating the high-weight content in fault coexistence and fault chains, and then retraining the model to adapt to the system evolution and new fault modes, analyze the actual maintenance data, identify and record how repairing a fault can prevent or mitigate the occurrence of other related faults; a specific implementation method is to extract detailed information on the actual fault handling from the fault maintenance management system, including the time of fault occurrence, faulty components, fault phenomena, handling measures and handling results, etc., to form a structured maintenance work order data table. Through data cleaning and data fusion, the maintenance work order data is associated with data such as fault diagnosis results and warning information to build a complete fault panorama data set; feature engineering and data mining are performed on the fault panorama data to extract key features such as fault mode, fault coexistence and fault chain; the Apriori algorithm is used to find association rules such as "the probability of fault A and fault B occurring at the same time is P" based on support and confidence indicators through recursive generation of frequent itemsets, and fault co-occurrence rules are mined. The apriori function in the mlxtend library of Python is used to set parameters such as minimum support and minimum confidence to control the number and quality of association rules. The PrefixSpan algorithm is used based on the idea of ​​prefix projection to recursively project the time series database onto a shorter prefix to generate frequent subsequence patterns, and a time series rule such as "the probability of fault B occurring x hours after fault A occurs is Q" is found. The pyprefixspan function is used to implement it. A knowledge graph is built based on the fault panorama data, and entities such as fault type, fault component, fault phenomenon, environmental factors and treatment measures are abstracted as nodes in the graph. , abstract the association relationship as the edge in the graph, store the knowledge graph in the graph database Neo4j, support incremental update, version management and query retrieval of the graph, use the entity relationship in the knowledge graph, adopt the random walk algorithm, start from the starting fault node, randomly select the next node according to the weight of the edge (such as the strength of fault correlation), and obtain the probability distribution of fault propagation after repeated many times. By setting the number of steps and repetitions of the random walk, the depth and coverage of the reasoning are controlled. The centrality index of the node is calculated using the degree_centrality and betweenness_centrality functions of the networkx library, and the centrality threshold is set to screen out the key nodes. The results of the random walk are combined with the key nodes to obtain a high-confidence fault chain to guide the maintenance priority sorting of key components. The fault maintenance feedback data is used as a new training sample. A sliding window method is used to maintain a fixed-size sample queue. Every time a new fault processing sample is received, it is added to the queue, and the oldest sample at the end of the queue is deleted. A decay weight is assigned to all samples in the queue, and an exponential decay function is used.That is, the weight of the sample decreases exponentially with its position in the queue, and then the weighted gradient descent method is used to update the model parameters. When calculating the gradient, the gradient of each sample is multiplied by its weight. By adjusting the parameters of the attenuation function, the importance of new samples is controlled, and the forgetting factor is introduced to regularly clear the sample queue to avoid the accumulation of outdated data, so as to achieve incremental update of the fault diagnosis model and adapt to the dynamic changes of the system operation conditions. The FP-Growth algorithm is used to mine fault prevention association rules, and the fault maintenance transaction data set is compressed by constructing an FP tree (frequent pattern tree) to avoid repeated scanning and generation of a large number of candidate sets by the Apriori algorithm. A transaction database is constructed based on the fault maintenance data, and each transaction contains a maintenance of a fault. The measure item set and the fault occurrence item set in the subsequent period of time are set with the minimum support and the minimum confidence, and the association rule of the form "maintenance measure X->fault prevention Y" is obtained. The support and promotion of the rules are sorted, and high-quality prevention strategies are screened out. Recommended suggestions for fault prevention are generated, and the FP-Growth algorithm in the mllib library of spark is used for distributed association rule mining; more specifically, the fault repair work order data table is cleaned through ETL tools such as Kettle, and the fields with more than 30% missing values ​​are removed. The numerical fields are normalized, and the text fields are segmented and feature hashed. The processed maintenance work orders and faults are analyzed using the Python Pandas library. The fault diagnosis and warning information data are associated and merged. The data with a timestamp error of less than 5 minutes is considered to point to the same fault. The Apriori algorithm is used for the fused fault panorama data. The support is set to 20%, the confidence is set to 60%, and the maximum frequent item set length is 3. The fault co-occurrence rules such as "inverter fault & high temperature warning -> photovoltaic panel damage" are mined. The PrefixSpan algorithm is used, and the minimum time support is set to 1 day. The fault timing mode such as "inverter high temperature warning -> (1 day later) inverter fault -> (3 days later) photovoltaic panel open circuit" is mined. The SF framework is used to build the knowledge graph and define EquipmentFailure and FailurePhenome Non and other 10 core ontology classes, define hasSymptom and leadTo and other 20 ontology relations, and use D2RQ tool to map business data in MySQL to ontology instances. Use Drools rule engine optimized based on RETE algorithm to convert the rules mined by Apriori and PrefixSpan into DRL format. In the random walk algorithm, the initial weight of each node is set to 1, and the next hop node is selected according to the edge weight in each iteration. After 1000 iterations, the steady-state distribution is obtained. The HITS algorithm of Gephi software is used to calculate the Hub and Authority centrality of the nodes in the fault propagation graph, and the top 10% of the nodes are taken as key nodes.Use Kafka's Logstash plug-in to collect incremental data from data sources such as MySQL and Neo4j in real time, aggregate it to HDFS through KafkaConnect, and then consume and filter it by SparkStreaming. The sliding window size is set to 30 days, the decay factor is set to 0.8, and the LightGBM classification model is trained to be updated in real time to predict the new fault diagnosis data. If the fault level prediction changes, an alarm is triggered. The FP-Growth algorithm is set to have a minimum support of 1%, a confidence of 10%, and a lift of 2. The mined fault prevention association rules are visualized in a word cloud, with an emphasis on rules with high Support, Confidence, and Lift indicators. Finally, the Airflow workflow engine is used to orchestrate various links such as data collection, cleaning, fusion, mining, reasoning, and service. A scheduled schedule is set at 2 a.m. every day. The average execution time of the Pipeline is controlled within 30 minutes, and the normal operating time exceeds. ,

[0101] As another example of this embodiment, according to the fault warning mechanism, by continuously monitoring the operating data and environmental factors of the photovoltaic system, the previously constructed AI-driven fault diagnosis model is applied to evaluate the system status in real time. Combined with the fault propagation path, the temperature of the photovoltaic panels, the status of the cable connections, and the voltage and current fluctuations of the inverter are continuously monitored. When signs of potential faults are detected, a warning signal is issued, and the fault type and location are provided, and its propagation path and impact range are analyzed.

[0102] A specific implementation method is to obtain the real-time operation data and environmental factor data of the photovoltaic system, including photovoltaic panel temperature, cable connection status, inverter voltage and current and other parameters, and perform data preprocessing and feature extraction. The preprocessed data is input into the constructed AI-driven fault diagnosis model. The model evaluates and predicts the photovoltaic system status in real time based on the input data. If the fault diagnosis model determines that there are signs of potential faults in the photovoltaic system, the fault warning mechanism is triggered, and the photovoltaic system automatically generates a fault warning signal. According to the diagnosis results of the fault diagnosis model, the potential fault type and location are determined, and presented to the operation and maintenance personnel in an intuitive way through data visualization technology. Graph theory is used to calculate the fault type and location of the photovoltaic system. The fault propagation path is analyzed by the method. By constructing a fault propagation diagram, other components or subsystems that may be affected by the fault are identified. If the fault propagation path analysis results show that the fault may affect key components, the operating parameters of the relevant components are automatically adjusted to reduce the impact of the fault. Through big data analysis technology, historical fault data and current fault characteristics are correlated and analyzed to determine the impact range of the current fault on the entire photovoltaic system. According to the impact range analysis results, a fault handling priority list is automatically generated and pushed to the operation and maintenance personnel to guide them to carry out targeted fault handling. During the fault handling process, the operating status of the photovoltaic system is continuously monitored, and real-time data feedback and fault The fault diagnosis model is re-evaluated and the fault handling strategy is dynamically adjusted until the fault is completely eliminated. Specifically, for example, real-time data such as the photovoltaic panel temperature of 42°C, the cable connection resistance of 3Ω, the inverter output voltage of 220V and the current of 8A are obtained, and the data is pre-processed by normalization and denoising. The 23 pre-processed feature parameters are input into the trained AI-driven fault diagnosis model. The model outputs a photovoltaic system status score of 92, indicating that the probability of potential faults is high. When the model predicts that the probability of photovoltaic system failure exceeds 85, a fault warning signal is automatically generated, and the operation and maintenance personnel are notified by SMS, email, etc. If the model diagnoses that the inverter IGBT module has a fault In the case of potential faults, the specific location coordinates (X:24, Y:13, Z:7) of the faulty components are marked on the system topology map through three-dimensional visualization technology. Through graph theory algorithm analysis, it is found that the inverter fault may propagate along the connecting line to the junction box and the distribution cabinet, and a fault propagation map is constructed. The fault propagation probabilities are 32 and 25 respectively. If the probability of the fault propagating to the distribution cabinet exceeds 3, the overvoltage protection threshold of the distribution cabinet is automatically adjusted from 500V to 450V to reduce the impact of the fault. Through big data analysis, it is found that the current fault characteristics are more than 85% similar to the characteristics of three inverter faults in history. It is judged that the current fault may affect three photovoltaic sub-arrays, accounting for 5% of the total installed capacity.According to the scope of fault impact, the fault handling priority is automatically generated, with the inverter fault ranked first (priority 95), followed by the combiner box (priority 85) and the distribution cabinet (priority 70). During the replacement of the inverter IGBT module, the PV system re-evaluates the inverter performance every 5 minutes. If the fault is not completely eliminated, the PV system automatically recommends repairing the inverter control board until the entire PV system resumes normal operation.

[0103] The embodiment of the present invention proposes a photovoltaic operation fault diagnosis method, which ensures the reliability of the data source by acquiring real-time monitoring data and historical record data, and provides a data basis for subsequent fault diagnosis; further improves the accuracy of the data by extracting features from the data and constructing feature vectors; obtains the association between different fault modes and identifies the key features of each fault mode by constructing a fault diagnosis model, so as to improve the accuracy of fault identification and diagnosis; and identifies and distinguishes various fault conditions in the system by labeling the fault feature vectors and constructing a label classification model, so as to improve the accuracy of fault analysis; and constructs a fault propagation model by simulating the system operation state and the fault propagation path, and obtains the fault. The data propagated from the source is used to analyze the propagation path of the fault and the potential impact of the fault on other components to improve the reliability of fault diagnosis. Finally, multiple models are constructed to diagnose and analyze the fault conditions of the system, generate effective fault diagnosis reports, and assist in taking preventive measures in advance to block the propagation of faults and ensure the stable operation of the system. The fault diagnosis data is integrated into the knowledge graph, the fault diagnosis model, the label classification model and the fault propagation model for feedback training, and the model is continuously updated to adapt to system evolution and newly discovered fault modes, including updating high-weight content in fault coexistence and fault chains, and then retraining the model to adapt to system evolution and new fault modes to ensure the reliability of system operation.

[0104] Example 2

[0105] See also Figure 2 , Figure 2 The module structure diagram of a photovoltaic operation fault diagnosis system provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, the embodiment of the present invention proposes a photovoltaic operation fault diagnosis system, including: a data construction module 201, a feature extraction module 202, a first model construction module 203, a data classification module 204, a second model construction module 205, a third model construction module 206 and a fault diagnosis module 207;

[0106] The data construction module 201 is used to construct a database of fault modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system;

[0107] The feature extraction module 202 is used to extract features from the fault mode and influencing factor database based on a preset feature extraction algorithm to obtain a fault feature vector;

[0108] The first model building module 203 is used to build a fault diagnosis model based on the fault feature vector through a preset decision algorithm;

[0109] The data classification module 204 is used to perform label classification on the fault feature vector based on a preset classification algorithm to obtain a label classification data set;

[0110] The second model building module 205 is used to build a label classification model based on the label classification data set;

[0111] The third model building module 206 is used to build a fault propagation model by simulating the system operation state and fault propagation path through a preset neural model based on the fault mode and influencing factor database;

[0112] The fault diagnosis module 207 is used to perform fault diagnosis analysis on the real-time monitoring data based on the fault diagnosis model, the label classification model and the fault propagation model to obtain a photovoltaic operation fault diagnosis report.

[0113] The embodiment of the present invention proposes a photovoltaic operation fault diagnosis system, which obtains real-time monitoring data and historical record data through a data construction module to ensure the reliability of the data source and provide a data basis for subsequent fault diagnosis; extracts features from the data through a feature extraction module to construct a feature vector to further improve the accuracy of the data; constructs a fault diagnosis model through a first model construction module to obtain the association between different fault modes and identify the key features of each fault mode, so as to improve the accuracy of fault identification and diagnosis; then, the fault feature vectors are respectively labeled and classified by a data classification module and a second model construction module to construct a label classification model, identify and distinguish various fault conditions in the system, so as to improve the accuracy of fault analysis; further, the system operation state and the fault propagation path are simulated through a third model construction module, a fault propagation model is constructed, data of the fault propagation from the source is obtained, the fault propagation path and the potential impact of the fault on other components are analyzed, so as to improve the reliability of fault diagnosis; finally, the fault diagnosis module uses multiple models to diagnose and analyze the fault conditions of the system, generates an effective fault diagnosis report, and assists in taking preventive measures in advance to block the propagation of faults and ensure the stable operation of the system.

[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

[0115] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0116] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

Claims

1. A photovoltaic operation fault diagnosis method, characterized in that: include: Build a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system; Based on a preset feature extraction algorithm, feature extraction is performed on the fault mode and influencing factor database to obtain a fault feature vector; Based on the fault feature vector, a fault diagnosis model is constructed through a preset decision algorithm; Based on a preset classification algorithm, label classification is performed on the fault feature vector to obtain a label classification data set; Based on the label classification data set, construct a label classification model; Based on the fault mode and influencing factor database, a fault propagation model is constructed by simulating the system operation state and fault propagation path through a preset neural model; Based on the fault diagnosis model, label classification model and fault propagation model, fault diagnosis and analysis are performed on the real-time monitoring data to obtain a photovoltaic operation fault diagnosis report.

2. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: The method of constructing a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system includes: Based on the preset deep learning algorithm, semantic extraction and several semantic features are performed on real-time monitoring data and historical record data; Based on a preset clustering algorithm, cluster analysis is performed on the plurality of semantic features to obtain different types of failure modes; Based on a preset association rule mining algorithm, association factor mining is performed on the historical record data and different types of failure modes to obtain an influencing factor database.

3. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: The method of extracting features from the fault mode and influencing factor database based on a preset feature extraction algorithm to obtain a fault feature vector includes: Based on a preset feature extraction algorithm, statistical features, trend features and periodic features of the failure mode and influencing factor database are extracted respectively to construct a feature group; The feature groups are combined to obtain a fault feature vector.

4. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: The method of constructing a fault diagnosis model based on the fault feature vector by a preset decision algorithm includes: Based on preset support and confidence thresholds, performing correlation analysis on the fault feature vector to obtain a correlation analysis result; Based on the correlation analysis result, the correlation between each feature vector and the fault mode is evaluated, and a key feature subset that meets the preset correlation requirement is selected; Based on a preset decision algorithm, a fault diagnosis model is constructed with the key feature subset as input and the relevant fault types as output.

5. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: The label classification is performed on the fault feature vector based on a preset classification algorithm to obtain a label classification data set, including: Based on the preset classification algorithm, the fault feature vector is labeled with multiple labels to obtain several multi-label fault feature vectors; By classifying the fault types of the plurality of multi-label fault feature vectors, a label classification data set is constructed.

6. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: Based on the label classification data set, a label classification model is constructed, including: An initial label classification model is constructed with the feature vector as input and the fault type as output; The initial label classification model is trained using the label classification data set to obtain a model training evaluation index; Based on the model training evaluation index, the model parameters and model structure of the initial label classification model are optimized until the preset performance requirements are met to obtain a label classification model.

7. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: Based on the fault mode and influencing factor database, a fault propagation model is constructed by simulating the system operation state and fault propagation path through a preset neural model, including: Inputting the fault mode and influencing factor database into a preset neural model for training to obtain simulation data of fault types and fault factors; Based on the simulation data of the fault type and the fault factor, simulate the system operation state and the fault propagation path to obtain the fault propagation simulation data; A fault propagation model is constructed based on the fault propagation simulation data.

8. A photovoltaic operation fault diagnosis method according to claim 1, characterized in that: Based on the fault diagnosis model, label classification model and fault propagation model, the real-time monitoring data is analyzed for fault diagnosis to obtain a photovoltaic operation fault diagnosis report, including: Based on the fault diagnosis model, the real-time monitoring data is subjected to fault trigger judgment to obtain a fault judgment result; Based on the label classification model, label classification is performed on the fault judgment result to obtain a fault classification result; Based on the fault propagation model, simulating the fault propagation path of the fault classification result to obtain a fault propagation result; Based on the fault judgment results, fault classification results and fault propagation results, a fault diagnosis report is output.

9. A photovoltaic operation fault diagnosis method according to any one of claims 1 to 8, characterized in that: The following steps are also included: Based on a preset knowledge graph framework, a number of fault diagnosis data are obtained by reading data from the fault diagnosis report; Performing weight configuration on the plurality of fault diagnosis data to obtain fault diagnosis feedback data; The fault diagnosis feedback data is input into the fault diagnosis model, the label classification model and the fault propagation model for feedback training to update the fault diagnosis model, the label classification model and the fault propagation model.

10. A photovoltaic operation fault diagnosis system, characterized in that: include: A data construction module, a feature extraction module, a first model construction module, a data classification module, a second model construction module, a third model construction module and a fault diagnosis module; The data construction module is used to construct a database of failure modes and influencing factors based on real-time monitoring data and historical record data of the photovoltaic system; The feature extraction module is used to extract features from the fault mode and influencing factor database based on a preset feature extraction algorithm to obtain a fault feature vector; The first model building module is used to build a fault diagnosis model based on the fault feature vector through a preset decision algorithm; The data classification module is used to perform label classification on the fault feature vector based on a preset classification algorithm to obtain a label classification data set; The second model building module is used to build a label classification model based on the label classification data set; The third model building module is used to build a fault propagation model by simulating the system operation state and fault propagation path through a preset neural model based on the fault mode and influencing factor database; The fault diagnosis module is used to perform fault diagnosis analysis on real-time monitoring data based on a fault diagnosis model, a label classification model and a fault propagation model to obtain a photovoltaic operation fault diagnosis report.

Images