DETECCIÓN DE FALLOS EN CONJUNTOS FOTOVOLTAICOS: UN ENFOQUE ROBUSTO Y REGULARIZADO DE APRENDIZAJE DE MÁQUINA

Abstract

In this paper, a robust data-driven method for fault detection in photovoltaic (PV) arrays is proposed. Our method is based on the random vector functional link networks (RVFLN) which has the advantage of randomly assigning hidden layer parameters with no tuning. To eliminate the effect of measurement noise and overfitting in the training process which reduce the fault detection accuracy, the sparse-regularization method is utilized which uses l2-norm with loss weighting factor to compute the output weights. To attain strong robustness against the outlier samples, the non-parametric kernel density estimation is employed to assign a loss weighting factor. Through rigorous simulation and experimental studies, we validate the performance of our proposed method in detecting the short and open circuit faults based on only the output current and voltage measurements of PV arrays. In addition to stronger robustness comparing with the least square-support vector machine, we also show that our proposed method provides 80% and 100% average detection accuracy for short circuit and open circuit, respectively. Key Words: Canonical correlation analysis, fault detection, photovoltaic array, random vector-link network, sparse regularization