The high penetration of renewable plants on the actual energy system brings to important benefits for the environment, but originates unexpected problems because of the production discontinuity. Accurate forecast of power production is essential for managing the electricity balance on the grid by the TSO. For the producer, the main application of a good forecast is probably the improvement of the offer on to the day ahead energy market. The main criticality in predicting PV power production by means of a NWP model is the forecast of the global irradiance on the PV cell, often tilted and with particular azimuth. The inclined global irradiance (GTI) is difficult to predict, because of the complexity in the evaluation of the direct and diffuse component contribution. In order to deal with these problems, a PV forecast system has been implemented in this study and tested on two Italian sites. Different forecast chains have been implemented, in order to test different global area models and various schemes for the prediction of the GTI on the basis of the global, diffuse and direct components. In order to reduce the forecast bias and to obtain the prediction of the plant production, an artificial neural network has been implemented, which is supplied by astronomical and meteorological inputs derived from the 0-72h ahead prediction by NWP models. The NN goodness has been tested depending on the different types of variables used in input and the different configurations of the NWP forecast chains on two experimental Italian sites.