Enhanced solar radiation nowcasting by machine-learning-based blending of data-driven, satellite-images and all-sky-imagers based models

The share of solar energy in the electricity systems of many countries in the world will reach unprecedented values in the coming decades, fostered by the mitigation of climate change and also by the economic competitiveness of this energy. Accurate solar radiation forecasting models are critical for the integration of the increasing solar energy in power systems.

In this work, the benefits obtained by blending seven models: four All-sky imagers (ASI)-based, two satellite images based (one using low resolution and other using high-resolution images) and a data-driven model, were analyzed. The use of two blending models (linear and Random Forest (RF)) and two blending approaches (General and Horizon) were explored. The horizon approach constructs a different blending model for each forecast horizon, while the general approach trains a single model valid for all horizons. The study is conducted in southern Spain and blending models provide one-minute resolution 90-minutes ahead GHI and DNI forecasts. Results show the General approach and the RF blending model to perform superior and to provide enhanced forecasts. The relative improvement in rRMSE obtained by model blending was up to 30% for GHI (40% for DNI), being maximum at lead times between 15 and 30 minutes and negligible at lead times greater than 50 minutes. Results also show that blending of just the data-driven model and the two satellite models (low and high resolution), without including the ASI-based models, performs similarly to those blending models that used as input the ASI-based models. Results then indicate that, for point nowcasting, the use of ASI-based forecasting systems can be avoided by using a suitable blending of data-driven, high resolution and low resolution satellite-images-based forecasting models.