Removing prediction lag in wave height forecasting using Neuro - Wavelet modeling technique

Forecasting of waves using ANN has been done by many researchers since last two decades in which use of previous wave heights is done for forecasting the same for few hours to few days in advance. These wave forecasting models exhibit lag in prediction timing which makes the univariate time series forecasting a futile attempt. This can be attributed to high autocorrelation between the last two observed wave heights. In the present work a hybrid technique called multilevel neuro-wavelet transform is used for forecasting significant wave heights up to 36 hr in advance at three locations around USA coastline using the previously measured SWHs at the same locations in order to remove the phase lag in prediction. The discrete wavelet transform (DWT) used in the present work for multiple times decomposes the time series into approximate (low) and detail (high) frequency components preventing any correlation between the sequentially observed wave heights. The neural network is then trained with these decorrelated approximate and detail wavelet coefficients. The outputs of networks during testing are reconstructed back using the inverse DWT. It was seen that the prediction lag in forecasting of significant wave height is completely removed by this hybrid multilevel neuro-wavelet technique. (C) 2014 Elsevier Ltd. All rights reserved.