Nonlinear Autoregressive Neural Networks for Air Temperature forecasting

In the field of climatic conditions forecasting, the linear classical time series models are inadequate for modelling and predicting accurately the air temperature variability. This work presents the novel application of nonlinear autoregressive neural networks (NAR) in air temperature forecasting. Hourly air temperature data were extracted from the Chania Airport station located at the island of Crete in Greece, for a 10-year period. The NAR is a multi-layer, dynamic, recurrent neural network that employs feedback connections for multi-horizon time series forecasting. Multiple NAR networks were trained with feedback connections for 6, 12 and 24 hours and for forecasting horizons up to 24-time steps. In the context of evaluating the performance of the trained NAR networks the Mean Absolute Error (MAE) was used and specifically the errors are examined in terms of their dependence with the atmospheric circulation. The results indicate that the use of a high degree of feedback decreases the forecasting error and increases the forecasting horizon of reliable air temperature estimation. The forecast error is dependent on the atmospheric circulation and higher MAE values are related with depressions that effect the study area.