Tracking Defective Panel on Photovoltaic Strings with Non-Intrusive Monitoring and Deep Learning

Photovoltaic (PV) generation systems are susceptible to various types of faults. Our objective is to identify unusual operating conditions in a photovoltaic string using only the voltage and current generated at its terminals. To achieve this, we collected voltage and current samples produced by a PV string consisting of six panels during typical operation and four possible types of faults: full panel shading, partial panel shading, panel short-circuits, and electrical arcs in conductor cables. The first three fault groups were further subdivided into six faults, one for each PV panel, resulting in 20 different operating conditions. We collected samples of the electrical system’s generation under each state in different climatic situations. We used the resulting dataset to train a convolutional neural network (CNN) and a classic machine learning method, K-Nearest Neighbors (KNN). The results showed that the CNN’s ability to learn the characteristics that identify each of the 20 operating conditions resulted in an average accuracy of 95.78%, while the KNN, taking into account previously defined features, achieved an accuracy of 86.34%.