River water level prediction in coastal catchment using hybridized relevance vector machine model with improved grasshopper optimization

Modelling river water level (WL) of a coastal catchment is much complex due to the tidal influences on river WL. A hybrid machine learning model based on relevance vector machine (RVM) and improved grasshopper optimization (IGOA) is proposed in this study for modelling hourly WL in a catchment located in the east coast of tropical peninsular Malaysia. Considering the non-linear relationship between inputs and output, a recursive elimination filter based on support vector machine (SVM-RFE) was employed for the selection of the best combination of inputs from antecedent WL and rainfall data for the prediction of WL one hour ahead. The performance of IGOA was compared with classical GOA and particle swarm optimization (PSO) algorithms. Besides, the performance of the hybrid RVM model was compared with the artificial neural network (ANN) models hybridized with the same optimization algorithms. The SVM-RFE selected 1-, 12- and 24-lags WL data and 1-lag rainfall data as the most potential inputs. The relative performance of the models revealed the reliability of RVM-IGOA in WL prediction of a coastal catchment. Significant improvement of model performance was noticed after optimization using IGOA with Nash-Sutcliff Efficiency (NSE) of 0.986 and 0.981, and Kling-Gupta Efficient (KGE) of 0.981 and 0.974 for RVM-IGOA and ANN-IGOA respectively, compared to the models hybridized using other optimization algorithms with NSE between 0.969 and 0.971, and KGE between 0.890 and 0.908. The study indicates the selection of predictors based on their non-linear relations with WL and better optimization of model parameters can improve model performance in simulation of highly complex hydrological phenomena like tidal river WL in a tropical coastal catchment.