An integrated framework of input determination for ensemble forecasts of monthly estuarine saltwater intrusion

Abstract

Mid- and long-term saltwater intrusion forecasts for estuaries are challenging due to a wide range of dynamic interactions and the limited amount of available data. This study proposes a tailor-made method for input determination for ensemble forecasts of monthly estuarine saltwater intrusion. The proposed method is based on determining the initial set of candidates by the combined use of Pearson’s Coefficient (r) and Maximal Information Coefficient (MIC); and afterwards reducing the dimension of the input data sets by Principal Component Analysis (PCA). The current study uses Bayesian Model Averaging (BMA) method to combine the forecasting results of Random Forest (RF), Support Vector Machine (SVM) and Elman Neural Network (ENN) models to create an integrated forecast. The proposed modeling approach was tested and compared with seven alternative procedures to forecast the monthly saltwater intrusion at the Pearl River Delta (PRD). The results indicated that: (a) the monthly dynamics of saltwater intrusion are more sensitive to the long-term solar activities than the local wind force; (b) the valuable non-linear signals hidden in the related time series could be identified by the combined use of r and MIC; (c) dynamic statistics related to low runoff, high antecedent chlorinity, strong tidal force, and strong wind force are preferable over the monthly average values as model inputs; and (d) the proposed method achieved highest forecast accuracy with Nash-Sutcliffe coefficient (NSE) of 0.78. This study provides insights to the input determination for data-driven models of complex estuarine saltwater intrusion.