A city-scale estimation of rooftop solar photovoltaic potential based on deep learning

Abstract

The estimation of rooftop solar photovoltaic (PV) potential is crucial for policymaking around sustainable energy plans. But it is difficult to accurately estimate the availability of rooftop area for solar radiation on a city-scale. In this study, a generic framework for estimating the rooftop solar PV potential on a city-scale using publicly available high-resolution satellite images is proposed. A deep learning-based method is developed to extract the rooftop area with image semantic segmentation automatically. A spatial optimization sampling strategy is developed to solve the labor-intensive problem when training the rooftop extraction model based on prior knowledge of urban and rural spatial layout and land use. In the case study of Nanjing, China, the labor cost on preparing the dataset for training the rooftop extraction model has been reduced by about 80% with the proposed spatial optimization sampling strategy. Meanwhile, the robustness of the rooftop extraction model in districts with different architectural styles and land use has been improved. The total rooftop area extracted was 330.36 km2, and the overall accuracy reached 0.92. The estimation results show that Nanjing has significant potential for rooftop-mounted PV installations, and the potential installed capacity reached 66 GW. The annual rooftop solar PV potential was approximately 311,853 GWh, with a corresponding estimated power generation of 49,897 GWh in 2019.