HEURISTIC METHODS FOR MODELLING THE BEHAVIOUR OF CLIMATE VARIABLES USING A MULTILAYER PERCEPTRON

Abstract An empirical methodology to fit a model for forecasting climate variables is developed. The method is based on an artificial neural network (ANN) of the multilayer perceptron type (MLP). To configure the model, the series of mean monthly minimum temperature (TminMean) data observed at a weather station in Avila (Central Spanish Plateau), belonging to the synoptic and climatological network of the Spanish National Institute of Meteorology (NIM), were used. Experiments were undertaken to determine the num ber of training and test patterns, the number of data per pattern, layer activation functions, the training algorithm, and the end-of-training condition that would allow later application of the model. We also experimentally determined the type of data pre-processing for which the model providedth e best yield Th. e following were considered: differentiation, deseasonalisation an, omalies, normalisation and standardisation. The resulting model was applied to the TminMean series at the stations of the synoptic and climatological network of the NIM on the Spanish Central Plateau. A high degree of fitting was observed between the real and simulated series, as shown by the values of the determination coefficients