A Fast Machine Learning Model for Large-Scale Estimation of Annual Solar Irradiation on Rooftops
Abstract
Rooftop-mounted solar photovoltaics have shown to be a promising technology to provide clean electricity in urban areas. Several large-scale studies have thus been conducted in different countries and cities worldwide to estimate their PV potential for the existing building stock using different methods. These methods, however, are time-consuming and computationally expensive. This paper introduces a Machine Learning approach to learn from existing datasets in order to estimate annual solar irradiation on building roofs (in kWh/m2) for large areas in a fast and computationally efficient manner. The estimation is based on input features extracted from digital surface models (roof tilt, roof aspect, roof shading) and annual global horizontal irradiation. Five ML models are compared, with Random Forests exhibiting the highest model accuracy. The model is trained using data of the Swiss Romandie area and is then applied to estimate annual rooftop solar irradiation in remaining Switzerland with an accuracy of 90.8%.