The reverse flow of seawater causes salinity in inland waterways and threatens water resources of the coastal population. In the Pearl River Delta, saltwater intrusion has resulted in a water crisis. In this study, we proposed a tailored approach to chlorinity prediction at complex delta systems like the Pearl River Delta. We identified the delayed predictors prior to optimization based on the maximal information coefficient (MIC) and Pearson's correlation coefficient (r). To achieve an ensemble simulation, a Bayesian model averaging (BMA) method was applied to integrate temporally sensitive empirical model predictions given by random forest (RF), support vector machine (SVM), and Elman neural network (ENN). The results showed that: (a) The ENN performed the worst among the three; (b) The BMA approach outperformed the individual models (i.e., RF, ENN, and SVM) in terms of Nash–Sutcliffe efficiency (NSE), and the percentage of bias (Pbias). The BMA weights reflect the model performance and the correlation of the predictions given by its ensemble models. (c) Our variable selection method resulted in a stronger model with greater interpretability.
This item's license is: Attribution-NonCommercial-NoDerivatives 4.0 International