Medium and short range numerical weather prediction models are unable to satisfy the astronomical constraints in terms of both temperature forecast accuracy and spatial resolution, although a local area model specifically designed for astronomical purposes could improve the performances of a standard model. From a theoretical point of view, an astronomical LAM could provide an actual RMS error below the 2 Celsius degrees: nowadays this is hard or impossible to accomplish. For the neural network model we have presented here, the extimated prediction capability shows an application threshold at 12 hours, being a limit to the application of such NN topology. Among the various attempts which can be carried out in order to improve the reported results in terms of both time prediction and accuracy, a new NN topology will be tested, where the medium-range forecasts provided by the ECMWF is used as an input node together with the local monitored time series. The new topology could in principle act as a post-processing engine for the ECMWF medium scale forecast, triggered by the local weather.
Acknowledgements
Franco Buffa kindly ackwnoledges the Regione Autonoma Sardegna for the financial support under contract ex art. 37 L.R. 2/94.
The Carlsberg Automated Meridian Circle telescope, which is operated jointly by Copenaghen University Observatory, the Royal Greenwich Observatory and the Real Instituto y Observatorio de la Armada en San Fernando, kindly provided us with the time series of meteorological parameters which have been used in this paper. The friendly support we got from Bob Argyle (RGO) is highly appreciated.
Dr. Marino Marroccu from CRS4, kindly discussed with us the Local Area Model topic.