Improved Time Series Forecasting by a Twofold Neural Predictor

The prediction of future values of environmental data sequences is mandatory to the cost-effective management of available resources. Consequently, the possibility to improve the prediction accuracy, even if for a little bit, could be a very important goal to be pursued. We propose, in the present paper, a twofold prediction scheme that can be used in this regard. It is based on the chaotic nature of the sequences to be predicted and on the consequent characteristics of their prediction errors. Such an approach is therefore suited to the prediction of environmental data sequences, which are often characterized by a chaotic behavior, eventually corrupted by noise. The improvement of the prediction accuracy is investigated by using a particular neural network as reference predictor, since it is already characterized by good performances in such cases. The results obtained by considering both synthetic and actual environmental data sequences, the latter concerning some pollution agents measured in the downtown of Rome, are encouraging in exploring further the proposed twofold prediction approach.