Feature Selection on Spatio-Temporal Data for Solar Irradiance Forecasting

Solar energy is currently among the most important and convenient renewable sources, with a great potential to reduce the use of fossil fuels. However, power generation from solar panels is very irregular and highly dependent on weather conditions. Therefore, solar irradiance forecasting is a fundamental task to ensure an efficient power management. In power plants, besides the temporal observations, it is essential to consider the spatial relationship between close photovoltaic panels. In this work, we study the importance of feature selection for forecasting solar irradiance time series using spatio-temporal data. The experimental study considers nine feature selection techniques and compares the predictive performance of four regression algorithms using the different subsets of features. The data used comes from two different locations in Canada with multiple solar panels. The results demonstrate that including the proper spatial information using feature selection, particularly the methods based on evolutionary computation, enhances significantly the forecasting accuracy.