Towards Decarbonization: Estimating Rooftops Photovoltaic Power Potential from Satellite Images and Solar Irradiation Data

Being an excellent source for producing clean energy, solar power is likely to become one of the world’s primary sources of energy. In geographical areas where the meteorological conditions are favorable for photovoltaic systems, solar energy can indeed provide not only for households’ energy demand, but also for average sized cities’ consumption. Unfortunately, despite the technological improvements and the increasing popularity of solar systems, estimating the photovoltaic potential of a roof has remained a time and resource consuming process which involves a manual evaluation of the site. With this thesis we aim at developing a data driven method based on Convolutional Neural Networks for estimating roofs’ photovoltaic power potential of a set of households randomly sampled across a region in central Italy, Lazio. While current state-of-the-art data driven methods for solar potential estimation rely on LIDAR data, available for very limited places in the world, our estimation is based on the analysis of widely, publicly available remote sensing images and solar irradiation data, making this approach applicable worldwide. Given a pair of geographic coordinates, our method downloads a satellite image depicting the rooftop of the building of interest, extracts the roof’s geometry, and determines the azimuth orientation of each of its planar segments. Leveraging solar irradiation data, it is then possible to provide an estimate of the rooftop photovoltaic potential at pixel-level. Whit this approach we demonstrate that Deep Learning Computer Vision techniques can provide reliable information for estimating PV power potential and assessing whether a roof is suitable to solar panels installation or not.