Water-Level Forecasting Using Neuro-wavelet Technique

Abstract

In the support of all ocean-related activities, it is necessary to predict the actual seawater levels as accurate as possible. The present work aims at forecasting the water levels from 3 to 6 weeks in advance at three locations: Dauphin Island, AL (Gulf of Mexico); Portland, ME (Gulf of Maine); and Cordova, AK (Gulf of Alaska) of divergent oceanic environment along the US coastline using neuro-wavelet technique (NWT) which is a combined approach of wavelet transform (WT) and artificial neural network. For this, time series of water-level anomaly (difference between the observed water level and harmonically predicted tidal level) was used to develop the NWT models at respective stations to predict the water levels for three different lead times from 3 to 6 weeks ahead. For this, hourly observed water levels along with harmonic tides were obtained from the National Oceanic and Atmospheric Administration of USA. The time series of water-level anomaly was decomposed using discrete wavelet transform (DWT) into low (approximate) and high (detail) frequency components. Further, these approximate coefficients were decomposed up to the desired level of decomposition (third and fifth levels) by multiresolution analysis of WT in order to provide more detailed and approximate components which ultimately provides relatively smooth varying amplitude series to develop the NWT models. Thus, the effect of autocorrelation in time series analysis was removed by decorrelating it using WT. Neural networks were trained with these decorrelated approximate and detailed wavelet coefficients. The outputs of networks during testing were reconstructed back using inverse DWT. Network-predicted anomaly was then added to harmonic tidal level to predict the water levels. Performance of NWT models was judged by drawing the water-level plots and other error measures. The NWT models performed reasonably well for all forecasting intervals at all the locations.