Photovoltaic modules fault detection, power output, and parameter estimation: A deep learning approach based on electroluminescence images

Accurately detecting faults in photovoltaic modules/cells and estimating their effective power output and parameters of the equivalent circuit representation of photovoltaic modules is becoming increasingly critical for both the reliability of associated systems and the efficiency of electricity production from renewable energy sources. Existing studies often work with datasets containing photovoltaic cells that exhibit one fault at a time, leading to the classification of photovoltaic cells with multiple faults as “mixed” faults. Moreover, factors such as cell alignment and specific fault types, collectively called “cell level features”, are not considered in current studies estimating the power output of a photovoltaic module. Therefore, this paper focuses on a comprehensive deep-learning pipeline to separately detect three types of faults in photovoltaic modules/cells using electroluminescence images. Furthermore, it addresses the estimation of the output power of photovoltaic modules and the series resistance of their equivalent circuit, considering the cell-level characteristics extracted from the electroluminescence images. The proposed model demonstrates its ability to detect “black core”, “crack”, and “edge” faults with global accuracies of 0.93, 0.868, and 0.95, respectively. Furthermore, the proposed model estimates the power output of photovoltaic modules with a normalized mean absolute error of 0.03547 and a normalized root mean squared error of 0.04892. This outperforms the base model that relies solely on non-pre-processed detected faults and significantly larger models adept at extracting features from the electroluminescence images. Moreover, the VGG16-based model estimates the series resistance in the equivalent circuit representation of photovoltaic modules with a normalized mean absolute error of 0.04472 and a normalized root mean squared error of 0.0622. • Defect detection can serve as a dimensionality reduction method. • Incorporating cell-level features improves power output prediction. • Larger models combined with detailed images result in better parameter prediction. • For detecting black-core and edge defects, an object detection approach is the best. • A classification approach works better for identifying cracks.