Photovoltaic power production forecasts with support vector regression: A study on the forecast horizon

The objective of this study is to verify how different forecast horizons affect the accuracy of a method to forecast photovoltaic power production using support vector regression and numerically predicted weather variables. One year of power production forecasts were done for a 1 MW photovoltaic power plant in Kitakyushu, Japan. Two forecast horizons were evaluated, up to 2 hours ahead and up to 25 hours ahead. The results showed a variation of the accuracy of the forecasts according to the forecast horizon. The root mean square error for 1 year of up-to-2-hours-ahead forecasts was 0.104 MWh; whereas for the up-to-25-hours-ahead it was 0.118 MWh. The mean absolute error was 0.065 MWh for the up-to-2-hours-ahead forecasts and 0.076 MWh for the up-to-25-hours-ahead forecasts. In percentage terms, the root mean square error and the mean absolute error increased 13% and 17%, respectively, with the increase of the forecast horizon.