Spatiotemporal feature encoded deep learning method for rooftop PV potential assessment

Abstract

Rooftop photovoltaic (PV) systems represent a promising solution for enhancing renewable energy utilization in urban landscapes. Accurate estimation of rooftop PV power generation potential is hindered by shading effects induced by complex urban morphology, which significantly reduce solar irradiance on rooftop surfaces and lead to prediction errors. Traditional shading simulation methods are computationally expensive, underscoring the need for a nuanced equilibrium between computational efficiency and assessment accuracy. In this study, we introduce an innovative deep learning framework that effectively encodes a diverse array of spatiotemporal data sources to accurately predict shadow casting and calculate rooftop PV potential. Specifically, utilizing physics-based ground truth, the incorporation of the U-Net network along with three-dimensional (3D) building specifics, solar resource data, and meteorological parameters enables us to make precise forecasts regarding temporal changes in rooftop shadow patterns. This not only enhances computational efficiency but also ensures a high level of precision in power generation predictions. Experimental assessments carried out in Futian District, Shenzhen, reveal that shading effects alone result in an average energy loss of 5.32 % across rooftops. Moreover, our framework demonstrates superior performance compared to physics-based models, achieving an average Mean Absolute Percentage Error (MAPE) of 2.85 % for annual energy generation potential and a mean Intersection over Union (mIoU) of 89.23 % for shading effect evaluation. In addition, the proposed framework achieves approximately 158× and 65× speedup over traditional ray-casting and optimized ray-tracing methods respectively, highlighting its strong suitability for large-scale urban energy evaluations. Our contributions encompass the development of a novel deep learning framework for rooftop PV potential assessment, enhanced computational efficiency in urban analyses, and a resilient generalization capability with high accuracy across various urban settings.