Accurate urban solar potential estimation empowered by multimodal 3-D building reconstruction: a case study in Landshut, Germany

Solar potential analysis is crucial in decision-making to fight against climate change. In the literature, it still remains a difficult task to calculate rooftop solar potential in complex urban areas, primarily due to the lack of precise building models. To address this issue, a novel workflow is proposed to first extract high-fidelity 3-D building models and then accurately estimate solar potential on buildings. A multimodal neural network is proposed to reconstruct detailed 3-D building models while leveraging data fusion of RGB images, digital surface models, and point clouds. Subsequently, the reconstructed buildings are used to estimate incoming solar insolation and to obtain detailed solar panel configurations. Building-scale potential direct current (DC) outputs are calculated using the estimated solar insolation and panel configurations. Comprehensive experiments and evaluations demonstrate the superiority of the proposed pipeline. Compared to other publicly available sources, the proposed method minimized the estimation errors – compared to manual annotations – of solar insolation and solar potential by a large margin. In the study area of Landshut in Germany, the residual of estimated solar insolation was reduced from 123.15 kWh/m2 to 52.36 kWh/m2, corresponding to an improvement of over 50 %. For the estimated total DC output originating from solar energy, a substantially lower error of 24.93 MWh was achieved, outperforming the baseline residual of 78.16 MWh. Through uncertainty and sensitivity analysis using Monte-Carlo simulations, the introduced method is proven to be statistically robust and produces reliable and realistic results that can be integrated into real-world practices. Finally, the potential alternating current output of Landshut was estimated to be approximately 370.05 GWh according to the conducted sensitivity analysis.