Comparison of strategies for multi-step ahead photovoltaic power forecasting models based on hybrid group method of data handling networks and least square support vector machine

The forecasting techniques are affected by the renewable sources randomness. Improvements of the prediction models with more accurate results and lower error are necessary for future development of the microgrids projects and of the economic dispatch sector. The LS-SVM (Least Square Support Vector Machine), a relatively unexplored neural network known as GMDH (Group Method of Data Handling) and a novel hybrid algorithm GLSSVM (Group Least Square Support Vector Machine), based on the combination of the first two models, were implemented to forecast the PV (Photovoltaic) output power at several time horizons up to 24 h. In order to improve the forecasting accuracy, each model was combined with three strategies for multi-step ahead forecast (Direct, Recursive and DirRec). A detail analysis of the normalized mean error is carried out to compare the different forecasting methods, using the historical PV output power data of a 960 kWP grid connected PV system in the south of Italy. The outcomes demonstrate the GLSSVM method with the DirRec strategy can give a normalized error of 2.92% under different weather conditions with evident improvements respect to the traditional ANN (Artificial Neural Network).