Salinity Time Series Prediction and Forecasting Using Dynamic Neural Networks in the Qiantang River Estuary

Abstract

The early warning of saltwater intrusion is an important work for ensuring the drinking water supplies. To forecast and predict the daily maximum salinity for the water withdrawn for the waterworks located along the Qiantang River, the nonlinear autoregressive networks with exogenous inputs (NARX) model was applied. Since the multivariate time series of flow, the tide range, the salinities and the water levels observed at 8 gauging stations have great impact on the salt concentration in the river, this will bring in a large number of inputs when these variables directly fed into the NARX model and add unnecessary model complexity and poor performance. Therefore, the dynamic principal component analysis (DPCA) was used to reduce the data redundancy. Simulation predicted results show that the NARX model using DPCA can predict salinity in the river accurately, moreover, this method not only reduces the input dimension and over-fit the equation, but also enhances the model performance and the generalization ability considerably.