A Hybrid Ensemble Learning Model for Short-Term Solar Irradiance Forecasting Using Historical Observations and Sky Images

A hybrid multi-modal ensemble learning model is proposed for short-term solar irradiance forecasting based on historical observations and sky images in this paper. In the proposed model, historical observations of solar irradiance are utilized to extract temporal characteristics, and ground-based sky images are used as an exogenous input to offer cloud cover information. A powerful ensemble learning model, Extreme Gradient Boosting (XGBoost), is employed to capture the function relationships between input features and future observations. Since mean squared error loss is sensitive to the extreme large or small historical irradiance, a novel loss function is proposed to improve robustness of XGBoost. In order to find out the best controlling parameters, Rao-1 algorithm is employed for its easy operation. To validate performance of the proposed method, a solar irradiance dataset containing three-year historical observations and ground-based sky images collected from Folsom is employed. Meanwhile, five commonly applied methods, LASSO, Ridge regression, support vector regression, boosted regression trees, the generic XGBoost, are considered as benchmarking methods. The forecasting horizons from 5 to 30 min are considered for all compared methods while two metrics, mean absolute error and root mean squared error, are computed. Experimental results prove that the proposed hybrid model has better forecasting performance compared with benchmarking methods over all forecasting horizons.