Enhancements in Day-Ahead Forecasts of Solar Irradiation with Machine Learning: A Novel Analysis with the Japanese Mesoscale Model

Abstract

AbstractThe objective of this study is to propose and evaluate a set of modifications to enhance a machine-learning-based method for forecasting day-ahead solar irradiation. To assess the proposed modifications, they were implemented in an initial forecast method, and their effectiveness was analyzed using two years of data on a national scale in Japan. In addition, the accuracy of the modified method was compared with one of the forecast methods for solar irradiation used by the Japan Meteorological Agency (JMA), namely, the mesoscale model (MSM). Such forecasts were made publicly available only recently, which makes this study one of the first ones to compare them with machine-learning-based forecasts. The annual root-mean-square error (RMSE) of local forecasts of the JMA-MSM varied from 0.1 to 0.14 kW h m−2; the regional equivalent varied from 0.062 to 0.091 kW h m−2. In comparison with these results, the modified model achieved an average RMSE reduction of 7.5% on the local scale and 16% on the regional scale. The modified model also had a skill score that was 23% higher than that of the JMA model. Furthermore, the performance of the JMA model had strong spatial and seasonal dependencies, which were reduced in the machine-learning-based forecasts. The results show that the proposed modifications are effective in reducing large forecasts errors, but they cannot compensate for situations in which the input data used to make the forecasts are highly inaccurate.