Growing of PV for electricity generation is one of the highest in the field of the renewable energies and this tendency is expected to continue in the next years. Due to the various seasonal, hourly and daily changes in climate, it is relatively difficult to find a suitable analytic model for predicting the performance of a grid-connected photovoltaic (GCPV) plant. In this paper, an artificial neural network is used for modelling and predicting the power produced by a 20 kW(p) GCPV plant installed on the roof top of the municipality of Trieste (latitude 45 degrees 40'N, longitude 13 degrees 46'E), Italy. An experimental database of climate (irradiance and air temperature) and electrical (power delivered to the grid) data from January 29th to May 25th 2009 has been used. Two ANN models have been developed and implemented on experimental climate and electrical data. The first one is a multivariate model based on the solar irradiance and the air temperature, while the second one is an univariate model which uses as input parameter only the solar irradiance. A database of 3437 patterns has been divided into two sets: the first (2989 patterns) is used for training the different ANN models, while the second (459 patterns) is used for testing and validating the proposed ANN models. Prediction performance measures such as correlation coefficient (r) and mean bias error (MBE) are presented. The results show that good effectiveness is obtained between the measured and predicted power produced by the 20 kW(p) GCPV plant. In fact, the found correlation coefficient is in the range 98-99%, while the mean bias error varies between 3.1% and 5.4%. (C) 2010 Elsevier Ltd. All rights reserved.