Towards city-scale urban rooftop PV potential assessment: a data-driven framework coupling solar radiation prediction and rooftop extraction

Urban rooftops represent an underutilized resource with considerable potential for photovoltaic (PV) deployment and urban carbon reduction. However, refined assessment of rooftop PV potential is often constrained by the limited availability of solar radiation data and insufficient rooftop extraction accuracy from remote sensing imagery. To address these issues, this study proposes a data-driven framework that couples a solar radiation prediction with building rooftop extraction. For solar radiation prediction, a support vector regression model optimized with the Grey Wolf Optimizer was developed using readily available meteorological variables. This design improves the applicability of solar radiation estimation in data-constrained regions. For rooftop extraction, an improved DeepLabV3 + model was developed by incorporating a hybrid Dice-cross-entropy loss, low-level feature enhancement, attention mechanisms, and transfer learning. The proposed framework was applied to Shenzhen city in China as a case study. The results show that the annual rooftop PV generation potential was equivalent to approximately 23.7% of the city’s total electricity consumption. This amount could fully meet residential electricity demand or supply about 52.7% of industrial electricity demand. These findings demonstrate the value of integrating solar radiation prediction and rooftop extraction for refined rooftop PV assessment in urban building clusters.