Abstract. In this paper we test different approaches to the statistical post-processing of gridded numerical surface air temperatures (provided by the European Centre for MediumRange Weather Forecasts) onto the temperature measured at surface weather stations located in the Italian region of Puglia. We consider simple post-processing techniques, like correction for altitude, linear regression from different input parameters and Kalman filtering, as well as a neural network training procedure, stabilised (i.e. driven into the absolute minimum of the error function over the learning set) by means of a Simulated Annealing method. A comparative analysis of the results shows that the performance with neural networks is the best. It is encouraging for systematic use in meteorological forecast-analysis service operations.