A neural network solution to the problem of frost prediction

Modelling and predicting meteorological behaviour using conventional algorithmic and statistical techniques has many problems. Models need to be constructed for a specific geography and climate, formulae suitable for one region will not in general be applicable to another region. Neural networks are nonparametric and can be trained to model the meteorological behaviour of any region for which there is sufficient data. In this paper, feed-forward neural networks are trained using data collected in Sicily between 1980 and 1983 to predict the occurrence of frost. The number and type of inputs was varied to give an indication of which were important to the performance of the neural predictor. The best network only failed to correctly predict frost on one occasion out of fifty unseen test days.