REDES NEURONALES PARA MODELAR PREDICCIÓN DE HELADAS

Abstract

A B S T R A C T In this work models based on neural networks of the backpropagation type were developed in order to predict the occurrence of frosts from meteorological data such as temperature, relative humidity, cloudiness and wind direction and speed. The training and the validation of the networks were made on the basis of 24 years of meteorological data corresponding to the Rio Cuarto station, Cordoba, Argentina. These data were grouped as follows: 10 years for the training data set and 14 years for the validation data set. Different models were built to evaluate the performance of the networks when different numbers of input variables and/or neurons in the hidden layer are used, and the probabilities of success in the prediction results on considering different input variables. In the models used, the percentage of days with prediction error was 2%, approximately, for the 14 years of application; when effective frosts days are considered the percentage varies between 10 and 23%, for the same period. The simulation results demonstrated the good performance and the relevance of this methodology for the estimation of the behavior of non-linear phenomena like frosts.