A compound of feature selection techniques to improve solar radiation forecasting

The Global Horizontal Solar Irradiance prediction (GHI) allows estimating in advance the future energy production of photovoltaic systems, thus ensuring their full integration into the electricity grids. This paper investigates the effectiveness of using exogenous inputs in performing short-term GHI forecasting. Thus, we identified a subset of relevant input variables for predicting GHI by applying different feature selection techniques. The results revealed that the most significant input variables for predicting GHI are ultraviolet index, cloud cover, air temperature, relative humidity, dew point, wind bearing, sunshine duration and hour-of-the-day. The predictive performance of the selected features was evaluated by feeding them into five different machine learning models based on Feedforward, Echo State, 1D-Convolutional, Long Short-Term Memory neural networks and Random Forest, respectively. Our Long Short-Term Memory solution presents the best prediction performance among the five models, predicting GHI up to 4 h ahead with a Mean Absolute Deviation (MAD) of 24.51%. Then, to demonstrate the effectiveness of using exogenous inputs for short-term GHI forecasting, we compare the multivariate models against their univariate counterparts. The results show that exogenous inputs significantly improve the forecasting performance for prediction horizons greater than 15 min, reducing errors by more than 22% in 4 h ahead predictions, while for very short prediction horizons (i.e. 15 min) the improvements are negligible.