A Machine Learning Methodology for Estimating Roof-top Photovoltaic Solar Energy Potential in Switzerland

Solar photovoltaic (PV) deployment on existing building roof-tops has proven to be one of the most viable large scale resources of sustainable energy for urban areas. While there have been many studies on roof-top integrated PV systems at building and neighborhood scale, estimating the PV energy potential through the available roof surface area in large scale, however, remains a challenge. This study proposes a methodology to estimate the roof-top solar PV potential for existing buildings at the commune level (the smallest administrative division) in Switzerland. In addition, while several studies suggest physical models to assess the photovoltaic solar energy potential, the current study proposes a computational data-based learning approach, using the following four steps: (1) monthly estimation of the main solar irradiance components (global horizontal, direct normal, diffuse horizontal) for the entire Switzerland, that is, the total amount of solar energy available from the Sun (2) processing of training and testing population and building data at a commune level, (3) estimation of the available roof-top surface (AR), average roof tilted angle (β), and average shading coefficient (S) for buildings in the urban areas, for each commune, (4) combination of the available roof-top surface area for each commune with the solar potential along with other parameters so as to estimate the actual solar PV potential. The first step is achieved through a Support Vector Regression (SVR), using solar irradiance satellite data with an average RMSE of 1.68 W/m2. In the second step, GIS have been used to aggregate the data at a commune level. In the third step, a supervised learning algorithm using SVR has been used so as to estimate AR, β and S in existing buildings, which can be extremely difficult data to obtain at a large scale. We use (i) population and building density data, (ii) land use data, and (iii) building typologies as input features. Aggregated values for AR, β and S for 46 communes in Geneva canton are used as a training output (labeled data) for the learning process. Finally, in the fourth step, by combining solar irradiance and building parameters predictions, we estimate the solar PV potential at the national scale.