Short-term solar power forecasting using Support Vector Regression and feed-forward NN

This work presents an off-line model, which can provide accurate photovoltaic forecasts for utility grid managers. For this purpose we assess the performance of two models, using well known supervised machine learning techniques, for intra-hour (15 min) solar power forecasting. The first model is created using Least Square Support Vector Regression (LS-SVR) and the second using Feed-forward Neural Network (FFNN) trained with Levenberg-Marquardt algorithm (LM). The two models are compared using five performance indicators, MAE, MSE, RMSE, R
²
and RRMSE (%). The best model shows a good results with an RRMSE of 15.23% and a coefficient of determination R
²
= 0.96.