AIR TEMPERATURE PREDICTION USING RECURRENT NEURAL MODELS WITH EMBEDDED TIME DELAYS

Dynamic Neural networks are appropriate models that may be used for time series prediction in the simulation of nonlinear processes since they make use of the effect of input values in previous time steps. The present study deals with recurrent neural networks with embedded time delays that are used for air temperature prediction at nine urban sites of the city of Athens. The sites are small urban sites with or without vegetation and include building verandas, courtyard, park, squares and streets. All these areas have almost different behavior to the temperature fluctuations so we used different models to succeed the best possible accuracy. The data used for the neural network's training, validation and testing were hourly values recorded at these areas during a hot whether summer period. Error statistics of the results showed a good fitting of the models. The values of MSE error were decreased at shaded places such as street with vegetation, the park and also the two verandas, the model being adequately fast. The results of the present study demonstrated that the number of hidden layer neurons ranging between 8 and 10 may give satisfactory results. In addition, by the increase of the number of delays in some cases, better results can be achieved, in terms of error statistics.