A Study on Machine Learning-Based Photovoltaic String Multifault Diagnosis Model

For the imbalanced distribution of photovoltaic component-level fault data samples, which are easily affected by environmental and internal factors, a machine learning-based multi-fault diagnosis strategy for photovoltaic string level is proposed. Firstly, the data of different fault states of photovoltaic components are cleaned, filtered, and the original fault features and higher-order fault features (RD, relative deviation) are constructed. Secondly, the prediction performance of three mainstream models, namely decision tree model (LightGBM), linear model (LR), and deep model (MLP) is compared, indicating that the performance of the decision tree model (LightGBM) is significantly better than the comparative models. Finally, the important features are using the SHAP technique, and the risk factors related to different fault types are analyzed from both the group and individual perspectives. The results show that the RD feature can effectively improve the performance of different models, with a 10% increase in AUC when comparing LightGBM models with and without RD. The decision tree model (LightGBM) is determined to be more suitable for diagnosing photovoltaic faults, with an AUC value of 0.939. The application of the SHAP technique provides more diversified explanations, analyzes the importance of data features for different fault types in both group and individual settings, and visualizes them, which will help provide decision references for on-site operation and maintenance personnel in fault analysis and handling.