Improved satellite-based intra-day solar forecasting with a chain of deep learning models


Satellite data and satellite-derived irradiance products have been extensively used in solar forecasting to better capture the spatio-temporal variations of solar irradiance. However, the potential advantages of using satellitederived irradiance and its improvements in solar forecasting have not been thoroughly explored. This work proposes a deep learning model chain with two models, one for deriving more accurate spatial global horizontal irradiance (GHI) estimates from satellite data, and the other for subsequently producing intra-day GHI forecasts using the derived spatial GHI. To evaluate the efficacy of the proposed method, GHI forecasts using different inputs are compared, namely, spectral satellite images (SAT), GHI estimates of the national solar radiation database (NSRDB), and satellite-derived GHI using deep learning (SAT-DL). The results show that satellitederived irradiance products (NSRDB and SAT-DL) generally outperform SAT. The improved GHI estimates of SAT-DL yield forecasts with lower normalized root mean square error (nRMSE), higher forecast skill, better ramp forecasts and forecast distributions, when compared with NSRDB for the cases studied. However, forecasting under frequent cloudy conditions is found to have enlarged nRMSE and compromised performance in ramp analysis, and forecasts are biased under high- and low-irradiance conditions. Despite these challenges, the deep learning model chain approach provides a novel framework for satellite-based solar forecasting that can yield more accurate forecasts than the benchmark deep learning methods, which is beneficial to a wide range of stakeholders in the solar energy sector.

Energy Conversion and Management