Data-driven prediction of fine-grained facade solar irradiance for urban PV potential assessment

Harnessing solar energy from building facades offers a significant opportunity for urban decarbonization, yet accurately assessing this potential is hindered by the computational cost of modeling complex shading in dense cityscapes. To address this precision-efficiency dilemma, we introduce SolarCViT, a novel, physics-informed deep learning (DL) framework that directly maps 3D urban geometry and solar conditions to fine-grained facade irradiance. By learning from high-fidelity, physics-based simulations, SolarCViT internalizes complex urban morphology and solar geometry, enabling rapid predictions without explicit ray-tracing during inference. The model achieves a Mean Absolute Error of 8.15 W/m2 and a Mean Relative Error of 4.63 %, while demonstrating a computational speedup of up to 48.23 times compared to physics-based simulations. We applied SolarCViT to evaluate the city-wide facade PV potential of Shenzhen, China, revealing an estimated annual generation capacity of 155,162 GWh. This capacity is equivalent to 137.5 % of the city’s 2023 electricity consumption and could offset up to 120.64 million metric tons of CO2 emissions annually. Our work demonstrates that a physics-informed DL approach can effectively resolve the trade-off between accuracy and scalability in city-scale solar assessment, providing a powerful and generalizable tool to guide urban energy transitions.