Evaluating IV curve derived features for fault detection

IV curves contain diagnostic information which characterizes faults in photovoltaic systems. Past research used IV curve derived features for fault detection, but a systematic investigation of features in outdoor conditions is missing. In this work, we perform outdoor IV measurements on module level with varying penetration of potential induced degradation, cell cracks, high series resistance and partial shade. We systematically evaluate eighteen IV derived features derived from the literature, investigating their ability to detect the faults. We calculate the feature’s importance for classifying the faults using machine learning classification methods (ExtraTrees, RandomForest, XGBoost, and GradientBoosting). Classification method-specific differences were observed and later eliminated by averaging the feature importance from all methods. Results show Rs, Vmp/Voc and FF were most important to detect high series resistance, Mppf, Vte and Imp/Isc for potential induced degradation, the FF, Imp/Isc and Vte for cell cracks, and Mppf, Vmp/Voc, Vmp and FF to detect shade at an overall classification accuracy of 95%. Greater importance was found for features that require IV and sensor based irradiance measurements compared to maximum power point monitoring. The method applies to any feature–fault combination and offers a strong indication of which features shall be further considered for fault detection strategies.