Abstract
Large-scale solar photovoltaic installations are required to have reliable inspection methodologies in order to detect surface conditions or degradation states which could negatively affect their operational efficiencies. In this paper, we develop a visual condition classification system for large-scale photovoltaic (PV) modules based on an AI technique capable of identifying six PV module surface conditions classified into six different classes: physically damaged, visually apparent electrical degraded, dirty/dusty, snow-covered, bird-drop impacted, and clean. Using transfer learning techniques, four different architectures were tested as classifiers under the same balanced classification protocol. EfficientNet-B0 produced the highest average results among all tested architectures with an average overall classification accuracy score of 94.3% for all tests. Additionally, the average Macro-Precision, Macro-Recall and Macro F1 scores for EfficientNet-B0 were 93.2%, 92.8%, and 93.0%. Furthermore, EfficientNet-B0 exhibited a variability in its Macro-Precision and Macro F1 scores of no greater than 1.6% between the best and worst test result of any individual run, thus demonstrating consistent performance across multiple runs. Lastly, the maximum processing time (inference time) for EfficientNet-B0 was less than 18 ms per image in the current reporting. The results demonstrated herein represent a proof-of-concept for a scalable methodology for visually inspecting PV surfaces, but additional testing at multiple sites along with the collection of additional data regarding operational status will be necessary before validation can be conducted.