Résumé:
his thesis undertook three main studies to detect and diagnose faults in photovoltaic systems. The first study involved simulating solar panels using a single-diode electrical circuit model and deriving the mathematical formula that characterizes the circuit's behavior.The GPC and EPC algorithms were used to determine the values of the five parameters in this model, and a database consisting of practical measurements was employed to assess the effectiveness of these algorithms. In the context of the second study, a novel method for fault detection and diagnosis in PV systems, based on the well-known GPC algorithm, was developed. This approach entails partitioning the training dataset into two hyper spheres, each representing a class, and only calculates the distances between a new data point and the center of each sphere. This eliminates the need to calculate distances across the entire dataset, as is required in classical KNN. In the last achievement of this thesis, another statistical algorithm for the detection and diagnosis of faults in photovoltaic systems was investigated. In contrast to the decision tree based on the Gini index, this algorithm computes Euclidean distances between a chosen point and the entire dataset. It extracts the minimum and maximum distances for each class, and arranging these distances in ascending order identifies one particular case among five. The faults are classified based on this identified case. To ensure the effective operation of both algorithms, four essential features are necessary: cell temperature, irradiance, as well as current and voltage at the maximum power point. Three distinct faults were taken into account: short circuit, open circuit, and partial shading. Finally, these methods were evaluated against a range of machine learning algorithms, such as SVM, DT, KNN, and RF. The obtained results using the developed algorithms demonstrated significant enhancements in accuracy, precision, and recall.
