The application of wavelets and artificial intelligence methods in hydrological forecasting (Q2888999)

The paper describes the application of wavelets and artificial neural networks (ANN) in order to obtain a one day flow prediction of rivers in regions of severe drought. The input to the prediction consists of flow data from the past; in particular the authors use periods of 6 and 8 days, respectively.NEWLINENEWLINEThe used ANN has a multilayer perceptron structure with one hidden layer and backpropagation for training. Wavelet preprocessing of the input data consists of a discrete wavelet transform (DWT) and the selection of certain components from the wavelet decomposition. This selection then serves as input to the ANN. The authors determine the prediction performance of the ANN by measuring the deviation predicted from actual data with standard procedures like the root mean square error (RMSE). Finally this performance is compared, when predicting with or without wavelet preprocessing. The authors report a good prediction performance of the ANN in general with slight improvement when wavelet preprocessing of the input data is done.NEWLINENEWLINEThe work is mostly experimental and touches wavelet topics only cursorily. Some details remain unclear, thus, e.g., no hints are given with respect to the used wavelet and the given formula for computing dyadic wavelet coefficients seems not to be very efficient as compared with standard filtering and subsampling procedures usually used in computing the DWT [\textit{S. Mallat}, A wavelet tour of signal processing. The sparse way. 3rd ed. Amsterdam: Elsevier/Academic Press (2009; Zbl 1170.94003)]. Moreover, when comparing wavelet preprocessing with raw data the authors do not comment on the reason for choosing for both cases different periods of past data for prediction. Nevertheless, the proposed method seems to work for one day prediction even if it remains open, whether wavelets really are responsible for that.