Deep Learning based Prediction of Solar Surface Irradiance with Geostationary Satellite Images

Solar energy has been expanding its scope for use in various systems, such as power generation, portable power, and power for eco-friendly moving objects and space power of various sizes depending on the purpose. However, since solar energy can be affected by meteorological and environmental variables, it is difficult to predict and manage energy production. Recently, artificial intelligence techniques have been applied to optimize predictive models and improve predictive performance of renewable solar energy. In this paper, a deep neural network-based prediction model is presented to predict the amount of solar energy potentials using geostationary satellite image data in units of one hour for over 7 years. In most of the previous studies, only a short-time prediction has been possible using only the daytime information, but in this model, the prediction performance is improved by using the image information including the cloud movement during the nighttime. In order to combine images of different characteristics of solar surface irradiance (SSI) and infrared (IR) in successive time and integrate them into one predictive model, after learning the basic structure to solve the trade-off problem between localization and context in the deep neural network structure, a generative model-based learning model is connected for matching between images of the different characteristics. In addition, in order to preserve the value of the target region that occupies a relatively small portion of the image, the performance of the model is supplemented using the region of interest (RoI) mask As a result, a model with improved predictive performance is presented when predicting using both daytime and nighttime image information the previous day.