Machine learning based fault detection in PV arrays

Renewable Energy Resources (RES) contribute significantly to sustainable energy generation given the increasing demand and global environmental degradation. Among various RES, solar energy is one of the most attractive power sources due to its technical, economic and environmental benefits. The increasing use of photovoltaic (PV) systems poses various technical problems previously unknown to the PV industry. PV array short circuit and open circuit faults are one of the problems that can cause PV failure and lower system efficiency. These faults are hardly detectable at low irradiance conditions or when the fault occurs between points with near electrical potential, i.e. low mismatch faults. Moreover, maximum power point tracking (MPPT) schemes can further increase the difficulty of detecting such faults while optimizing the power output of a PV array under different operating conditions. Due to the stated reasons, it is vital to detect such faults in order for the PV system to operate stably and with maximum efficiency. In this thesis, a robust data-driven method is proposed for fault detection in PV arrays. Our method is based on random vector functional link networks (RVFLN) that have the advantage of randomly assigning hidden layer parameters without adjustment. The sparse regulation method using l2-norm with loss weight factor was used to eliminate the effects such as measurement noise, long training time and overfitting, which reduce fault detection accuracy, and to calculate output weights. To obtain a strong robustness against outliers samples, the non-parametric kernel density estimation was used to assign a loss weighting factor. Through simulation and experimental studies, the performance of our proposed method has been shown to be successful in detecting short and open circuit faults based solely on the output current and voltage measurements of PV arrays. In addition to the stronger robustness compared to the smallest square support vector machine, we have also shown that the method we propose provides an average detection accuracy of 80% and 100% for short circuit and open circuit, respectively.