Local and Regional Hour-Ahead Forecasts of Solar Irradiance with Training Data Selection and Support Vector Regression

抄録

In markets with high penetration of photovoltaic power, methods to forecasts of solar irradiance one hour ahead of time are expected to provide useful information to execute services of secondary and tertiary regulation of power systems load. The objective of this study is to propose a method to forecast solar irradiance, one hour ahead of time, using numerical weather prediction and recently measured data. The proposed method uses a support vector regression algorithm with a training data selection approach to yield the best possible forecasts for each hour. We verify the validity of the proposed method using it to forecast one year of hourly solar irradiance in local and regional scale for the Kanto region in Japan. For the local forecasts, the method yielded forecast root mean square errors of 0.060 to 0.065kWh/m² and mean absolute errors ranging from 0.031 to 0.034kWh/m². These errors were calculated with data from 5h to 20h of each target day. In regional scale, both types of errors were reduced to 0.032kWh/m² and to 0.019kWh/m², respectively. Finally, regardless the spatial scale used, the forecasts of the proposed method outperformed considerably reference forecasts based on persistence. Local and regional skills scores varied between 0.67 to 0.73 for the former, and 0.97 for the regional case. These results show indicate the good performance of the proposed method.