An Original Contribution to Photovoltaic Cell Defect Characterisation Using Electroluminescence Imaging and Convolutional Neural Network Analysis for Nigerian Solar Installations

This research develops an original methodology for photovoltaic cell defect characterisation combining electroluminescence imaging with custom convolutional neural network architectures trained on a novel Nigerian solar installation defect dataset, advancing the state of the art in non-destructive PV module quality assessment for tropical operating environments. PV module defects including micro-cracks, cell interconnect breaks, potential-induced degradation, and localised shunting progressively reduce energy yield and can cause hot-spot fires in large-scale installations, necessitating systematic inspection programmes. However, existing EL image analysis CNN models are trained on datasets collected in European and East Asian climates and exhibit reduced accuracy on Nigerian solar module defect patterns, which are influenced by the distinct thermal cycling regime, Harmattan dust mechanical stress, and the specific module construction employed in Nigerian market products. The original contributions include the creation of the first annotated EL image dataset of PV module defects collected from Nigerian solar installations, comprising 4,200 images from 18 ground-mounted and rooftop systems in six states spanning all geopolitical zones; the development of a novel CNN architecture incorporating defect morphology priors derived from thermomechanical finite element analysis of module stress distributions under Nigerian thermal profiles; and the derivation of a theoretical relationship between EL image defect signature characteristics and the resulting power output degradation magnitude, enabling quantitative yield impact estimation directly from EL inspection results. The model achieves 96.1 percent defect classification accuracy on the Nigerian test set, enabling cost-effective automated inspection using drone-mounted EL cameras. Keywords: photovoltaic, electroluminescence, defect characterisation, convolutional neural network, Nigeria.