The purpose of this work is to predict the power produced by all the wind farms located over the entire area of Sicily island (one of the Italian market regions). The study has been conducted for a two-year period, considering hourly data of the aggregated wind power output of the island. For each day of the study and for the time horizons from 0 to 72 hours ahead, wind fields have been forecasted using a mesoscale meteorological model (RAMS) and ECMWF determinist forecast fields as boundary conditions. A Principal Component Analysis (PCA) has been applied on wind speed and wind direction data extracted at 50 m above ground on the model grid points inside the Sicily territory. A Neural Network (NN) has been then used as post processing technique of the PCA output to obtain the final wind power forecast. The input of the NN, for every forecast lead time, are the final forecast data of wind speed and direction after the PCA application. The NN has been trained both with the PCA output and the power measurements on the first year of the analyzed period. For sake of comparison an alternative approach, applying NN directly on RAMS output (without PCA), has been adopted too. The study shows that the PCA introduction leads to better results in terms of RMSE, MAE and BIAS and to a lower computational time.