Deep learning and explainable Artificial Intelligence for large-scale photovoltaic suitability analysis in Australian cities
Jie Shen, Fanghao Zheng, Fiseha B. Tesema, Wu Deng, Anthony Graham Bellotti, Jing Xie
Abstract
To support the achievement of the United Nations Sustainable Development Goals (SDGs), the global energy sector is undergoing a significant transformation towards renewable sources, with solar PV technology at the forefront. In 2023, global installed PV capacity surpassed 1.6 terawatts (TW), reflecting an accelerating trend in renewable energy adoption. Consequently, selecting suitable sites for PV development and analyzing PV suitability have become increasingly critical. A large body of research focuses on this area, reflecting the varying priorities of different countries or regions regarding PV suitability. The rapid expansion of PV in Australia provides an opportunity to reevaluate whether solar radiation should be considered the primary factor. This study investigates the key drivers of PV suitability in urban areas by assuming that existing PV installations reflect realistic locational conditions. High-resolution spatial datasets incorporating geographic, environmental, social, and urban planning variables were integrated with several deep learning models, U-Net, ResNet, CNN+Transformer, FCN, and SimpleCNN, as well as the AHP for comparative analysis. Results indicate that Res-Net achieves the highest accuracy in predicting PV distribution. XAI analysis further reveals that the top 2 influential factors across cities are Population Distribution (8.96%) and Forest Coverage (8.09%), while GHI ranks only third (7.48%). These findings uncover the intrinsic logic within the actual PV distribution and suggest that in urban environments of Australia, social and infrastructural variables may outweigh pure solar-related indicators. This study offers new insights for optimizing PV planning strategies in diverse global contexts.