The impact of training period for forecasting clearness index and insolation using support vector regression

Forecasting insolation of the next day is expected as a method to deal with photovoltaic power fluctuation. In general, extending the training period as long as possible is desirable for the training of a machine learning. However, annual variation of global horizontal insolation could affect the accuracy of insolation forecast. Thus, the objective of this study is to forecast clearness index and to investigating the forecast error of forecast models with various training periods. The proposed clearness index forecast model uses temperature, absolute humidity, cloudiness at each level from GPV-MSM, and air mass as input data into support vector regression. Also, the model was divided into three models depending on the solar height. Results showed that root mean square error, mean absolute error, and mean bias error were better compared to the insolation forecast. Consequently, the clearness index forecast model is preferable to eliminate a negative effect of annual variation.