Accurate water level predictions in a tidal reach: Integration of Physics-based and Machine learning approaches

Abstract

Accurate water level prediction is very important for coastal construction and flood prevention in an estuarine area. However, it is challenging to represent the water level processes accurately in a spatial domain due to the dual influences of river discharge from the upstream tidal reach (upstream discharge) and astronomical tides. In this study, a new method is proposed to predict water levels in tidal reaches by developing a riverine-estuary Hydrologic-Hydrodynamic Coupling model (H2C) and integrating the outputs (Upstream discharge and water levels, tidal levels) with long-short term memory. Then, the relative contribution of upstream discharge to water level predictions is assessed by extreme gradient boosting. The coupling model (H2C-XL) is driven by basin-scale Satellite-derived Meteorology Estimate datasets and TPXO9 tidal data. H2C-XL is tested in a tidal reach (Tianhe-Zhuyin reach) and found to be effective in enhancing the accuracy of water level predictions dramatically. The Nash coefficient and Kling-Gupta Efficiency values of predicted upstream discharge are 0.866 and 0.922, respectively. Accurate upstream discharge prediction is crucial for water level predictions in tidal reaches. This is reflected in that the Nash coefficient and Kling-Gupta Efficiency values are up to 0.781 and 0.806 for predicting water levels at upper hydrologic stations (Jiangmen and Daao stations), which are improved by 12.34% − 40.46% and 16.98% − 32.34% than those of above models. The relative contribution of upstream discharge to water level predictions in Tianhe-Zhuyin reach ranges from 0.37 to 0.55. The values vary at the same site depending on the hydrologic conditions each year and are relatively low near the coastline, but nonlinearly increase with the distance from the coastline.