Estimating the spatial distribution of solar photovoltaic power generation potential on different types of rural rooftops using a deep learning network applied to satellite images

Abstract

Rooftop photovoltaic (PV) power generation is an important form of solar energy development, especially in rural areas where there is a large quantity of idle rural building roofs. Existing methods to estimate the spatial distribution of PV power generation potential are either unable to obtain spatial information or are too expensive to be applied in rural areas. Herein, we propose a novel approach to estimate the spatial distribution of the general potential of rural rooftop power from publicly available satellite images. We divide rural building roofs into three categories based on their orientation and roof angle and propose a revised U-Net deep learning network to extract roof images from satellite images at the macro level. Based on the rooftop detection, a calculation method for the potential area of the installed PV panel at the micro level was developed, considering different types of PV panels and their maintenance methods. By combining the above results and setting the solar radiation parameters and PV system efficiency, we can obtain the spatial distribution of the rooftop PV power generation potential in rural areas. This method is applied in northern China on a village and a town scale, and the overall accuracy of the revised U-Net model can reach over 92%. The spatial distribution information was analyzed and displayed. The annual average PV power generation potential ranges from 26.5 to 36.2 MWh per household and from 7.3 to 10 GWh per village.