MLGA-GNN: Assessing horizontal and vertical heterogeneity in the photovoltaic potential of urban-scale building facades

Assessing the photovoltaic (PV) potential of urban building facades is crucial for achieving urban energy transition and carbon reduction goals. However, traditional assessment methods overlook the vertical variation of solar radiation intensity (SRI) on building facades, potentially underestimating the PV potential of local areas. This study proposes a novel method that vertically divides building facades into layers and predicts the solar radiation intensity (SRI) of each layer to evaluate the PV potential of building facades. Taking New York City as a case study, our results show that the proposed Multi-Layer Geo-Attention Graph Neural Network (MLGA-GNN) achieves better prediction performance than traditional machine learning and deep learning models. The spatial distribution of SRI shows lower values in dense, high-rise city centers due to shading effects, and higher values in peripheral areas with shorter, more spaced buildings. In the vertical direction, the SRI increases with height, forming a clear upward trend along with building elevations. Further analysis reveals that the MLGA-GNN can uncover significant differences in SRI between the top and bottom layers of buildings, helping to identify local high-irradiance areas overlooked by traditional assessment methods. Moreover, economic benefit and carbon reduction analyses based on different investment payback periods indicate that a 10–15-year payback period achieves a good balance between returns and costs. This research enriches the methods for assessing the PV potential of buildings at an urban scale, and provides decision support for the macro-layout and micro-design of urban PV systems.