A Hybrid Methodology for Salinity Time Series Forecasting Based on Wavelet Transform and NARX Neural Networks

Abstract

The forecasting and prediction of saltwater intrusion for waterworks located along the Qiantang River is an important function for ensuring the safety of drinking water supplies. However, for water resources, the majority of hydrology time series are non-stationary signals. A hybrid multivariable model using wavelet transform and the nonlinear autoregressive networks with exogenous inputs (NARX) was developed for the simulation and prediction of the daily maximum salinity time series. In this model, the non-stationary salinity time series and the factors which affect salinity such as daily flow, tide range and water level are decomposed into approximate sub-signal and several detailed sub-signals using the maximal overlap discrete wavelet transform-based multi-resolution analysis, and these decomposed sub-signals are then modeled and predicted using NARX neural networks, Finally, the predicted sub-signals are synthesized to obtain the predicted salinity time series. Taking the daily maximum salinity at the Qibao hydrological station on the Qiantang River as an example, the daily salinity was simulated by using the NARX, wavelet-NARX, wavelet-multiple Elman and wavelet-multiple NARX (WT-mNARX) models proposed in this article. It is shown that the WT-mNARX model has a higher prediction accuracy and better generalization capability than the other models, and, in this example, the NARX neural networks is very suitable for time series modeling.