Fast and accurate estimation of solar irradiation on building rooftops in Hong Kong: A machine learning-based parameterization approach

Abstract

Harvesting solar energy on rooftops can be a promising solution for providing affordable energy. This requires accurately estimating spatio-temporal solar photovoltaic (PV) potential on urban surfaces. However, it is still a challenge to obtain a fast and accurate estimation of rooftop solar PV potential over large urban built-up areas. Thus, this study proposes a parametric-based method to estimate annual rooftop solar irradiation at a fine spatial resolution. Specifically, seven parameters (Digital Surface Model, Sky View Factor, shadow from buildings, shadow from terrain, building volume to façade ratio, slope, and aspect) are determined that having great importance in modeling rooftop solar irradiation. Three machine learning methods (Random Forest (RF), Gradient Boost Regression Tree (GBRT), and AdaBoost) trained by the selected parameters are cross-compared based on R2, Mean Absolute Error (MAE), and computation time. As a case study in Hong Kong, China, the RF outperformed GBRT and AdaBoost, with R2=0.77 and MAE=22.83kWh/m2/year. The time for training and prediction of rooftop solar irradiation is within 13 h, achieving a 99.32% reduction in time compared to the physical-based hemispherical viewshed algorithm. These results suggest that the proposed method can provide an accurate and fast estimation of rooftop solar irradiation for large datasets.