Wavelet neural network processing of urban air pollution

We present a multi-resolution dynamic forecasting system (MDFS) based on neural networks for multi-step prediction of a time series of urban air pollutant data (hydrocarbons, HC). The MDFS utilizes the wavelet transform and the Daubechies mother wavelet to compute the wavelet coefficients of the original signal at various scales and a recurrent neural network (RNN) in the wavelet coefficient space to form a set of dynamic non-linear models of the sub-bands of the data. The decomposition strategy is suggested by the Fourier analysis of the time series showing cyclical components. The global system is capable of predicting the time series of the HC data with both long-term (coarse) and short-term (fine) resolution. The proposed approach can manage nonstationarity in the data and is suitable for online computation