DEM-based convolutional neural network modeling for estimation of solar irradiation: Comparison of the effect of DEM resolutions

Abstract

Recently, the use of solar panels has increased because of the growing demand for solar energy. To determine appropriate installation sites of photovoltaic (PV) panels, estimating the available solar energy at a certain area is important to predict the amount of power generation and planning PV plant operations. However, traditional data-driven approaches (e.g., machine learning) do not fully reflect the topographical characteristics of surrounding regions in the solar energy estimation, and the impact of data resolution (e.g., map scales) on the prediction accuracy has rarely been investigated. Thus, this paper presents a solar irradiation prediction model using a convolutional neural network (CNN) designed to process digital elevation map (DEM) images. Furthermore, an analysis of the impact of two different resolutions (i.e., 30 m and 60 m resolutions) on the model performance is also presented. A total of 25,000 DEM images are extracted from the national map of South Korea for both resolutions and then used as an input to train the CNN models. The results show that the CNN-based prediction models can be used to estimate the solar irradiation with high accuracy (e.g., mean square errors of 0.0018 and 0.0032 for 30 m and 60 m resolutions). It was also found that data resolution affects the performance of the CNN-based models. With an accurate estimation of the available solar energy at a certain site, the sites generating more power can potentially be evaluated and selected by searching a DEM on a large scale.