Deep learning methods for intra-day cloudiness prediction using geostationary satellite images in a solar forecasting framework

Abstract

Accurate solar resource forecasting remains a challenge. Electricity grid applications require both days-ahead and intra-day prediction. Satellite-based methods are known to be the best option for hourly intra-day solar forecasts up to some hours ahead. An adapted Deep Learning (DL) method has been recently reported to outperform the traditional Cloud Motion Vectors (CMV) strategy. This article analyzes the utilization of a well-documented computer vision DL architecture, the U-Net in various forms, for the satellite Earth albedo forecast problem (cloudiness), a straightforward proxy for solar irradiance forecast. It is shown that the U-Net performs better than advanced and optimized CMV techniques and previous art IrradianceNet, setting it at the state-of-the-art. The tests are done over the Pampa Húmeda region of southeast South America, an area in which challenging cloud conditions are frequent. The data for this study are GOES-16 visible channel images. These images present a finer spatial (≃1 km/pixel) and temporal (10 min) resolution than previously explored data sources for solar forecasting. Moreover, the image size used here is ×4 bigger (1024 × 1024 pixels) and the predictions reach further into the future (5 h) than in previous works. The analysis includes several ablation studies, involving different architectures, optimization objectives, inputs, and network sizes. The U-Net is optimized for direct and differential image prediction, being the latter a better-performing option. More notably, the U-Net models are shown to be able to predict cloud extinction, something that has been a barrier for CMV methods.