A current perspective on the accuracy of incoming solar energy forecasting

Abstract

Abstract The state-of-the-art in the accuracy of solar resources forecasting is obtained by using results reported in 1705 accuracy tests reported in several geographic regions (North America, Europe, Asia, Australia). There is a tendency in literature to avoid reporting model performance assessments in places, or on time-horizons, where it is known that the models have low performance. Usual bias and spreading statistical indicators are useful tools but taking proper decision by managers of funding agencies and policy makers requires usage of additional error measures. Several classes of forecasting models have been compared: persistence, classical statistics, machine learning, cloud-motion tracking, numerical weather prediction and hybrid models (combining classical statistics with machine learning approach and/or exogenous inputs). Forecasting errors increase by increasing the time-horizon. Forecasting model performance depends on time-horizon and climate. Machine learning and hybrid models have the best performance for intra-hour performance in all climates. The best intra-hour forecasts are provided in all climates by machine learning and classical statistics models while the hybrid models perform well in tropical and snow climates. For day-ahead forecasts the hybrid models have very good performance in all climates. Generally, the hybrid models have the best performance. The accuracy of the forecasting models significantly increased in the last decade. The normalized MBE and RMSE have been reduced by two thirds and one third, respectively.