Enhancing solar PV output forecast by integrating ground and satellite observations with deep learning

Abstract

Accurate output forecasts are essential for photovoltaic projects to achieve stable power supply. Traditional forecasts based on ground observation time series are widely troubled by the phase lag issue due to the incomplete consideration of the impacts of cloud motion. With the consensus that this issue can be addressed by introducing satellite-derived cloud information, we propose an innovative framework that integrates ground and satellite observations through deep learning to enhance PV output forecasts. Cloud motion patterns are captured from satellite observations using convolutional neural networks, and the long-range spatio-temporal cloud impacts on subsequent PV outputs are established by long short-term memory network. The forecast accuracy of real-time PV output is significantly improved, with a minimum (maximum) relative root mean square error of 16% (29%). The ratio of phase lag is reduced to 15% on average. This work provides a potential for alleviating the power intermittency of solar PV system and making advance planning in solar energy utilization.