A major challenge of the next years in global development will be the large scale introduction of renewable non-programmable (wind and solar) energy sources into existing energy supply structures. Due mainly to the variability of weather and shadow conditions, the total power production coming from photovoltaic plants in a specified future time period cannot be determined precisely, as it is a nondeter-ministic and stochastic process. This instability is caused by the dependence of PV generation on meteorological conditions: irradiance and temperature. If meteorological conditions can be forecasted with sufficient precision, it will be possible to estimate the energy a PV system will produce, making photovoltaic a more reliable electricity source. For this reason, in this paper, forecast solar radiation data, provided by two weather providers, have been analyzed. In order to verify their effectiveness, these forecast data have been compared with the measured ones and the errors have been calculated by means of the normalized Root Mean Square Error (nRMSE). Then an algorithm, that allows to classify a day as variable, cloudy, slightly cloudy or clear, has been implemented. Based on this classification, a maximum forecast error is determined. In this context, a neural network has been implemented, it allows to predict the nRMSE of a specific day knowing the percentages of the variable, cloudy, slightly cloudy or clear intervals (considered in that day) calculated on forecast data. Referring to Catania (Italy), experimental data are reported to demonstrate the potentiality of the adopted solutions. (C) 2012 Elsevier Ltd. All rights reserved.