On the Falult Diagnosis Method of Photovltaic Arrays Based on DCS-XGBoost

To address the issue of photovoltaic array fault diagnosis with small sample training sets, a Differentiated Creative Search (DCS) algorithm is proposed to optimize the hyperparameter configuration of eXtreme Gradient Boosting (XGBoost). Firstly, after establishing a photovoltaic array fault model, we simulate the model to obtain the fault dataset. The 11 feature vectors as inputs to the diagnostic model are constructed, which is a prerequisite for the subsequent design of a DCS-XGBoost-based photovoltaic array fault diagnosis algorithm.The algorithm is then compared with decision tree algorithms, support vector machines (SVM), and backpropagation neural networks (BPNN) using only $9 \%$ of the training set for validation. Secondly, the DCS algorithm, Crown Hedgehog Optimization Algorithm (CPO), Newton-Raphson Optimization Algorithm (NRBO), Kepler Optimization Algorithm (KOA), Elephant Herding Optimization Algorithm (EHO), and PID Search Algorithm (PSA) are employed to optimize the hyperparameters of the XGBoost model for classifying different types of faults. Comparative experiments over several metaheuristic-optimized XGBoost confirm that DCS-XGBoost reaches $99.72 \%$ classification accuracy with just $9 \%$ of the training set, enabling effective diagnosis of single and mixed faults. In addition, it demonstrates superior to overall performance with regard to accuracy, generalization, reliability, and training efficiency.