Coupling and Comparison of Physical Mechanism and Machine Learning Models for Water Level Simulation in Plain River Network Area

In this study, the JiaoGang Basin in the Yangtze River Delta plains of the river network area was the research object. A basin water level simulation model was constructed based on the physical mechanism model and Mike software, and the parameters were calibrated and validated. Based on the dataset produced by the physical model, three types of ML models, Support Vector Machine (SVM), random forest (RF), and gradient boosting decision tree (GBDT), were constructed, trained, validated, and compared with the physical model. The results showed that the physical mechanism model met the water level simulation accuracy requirements at most stations. In the training and validation periods, the RF water level simulation and GBDT water level simulation models had root mean square errors (RMSEs) of all stations less than 0.25 and the Nash–Sutcliffe coefficient (NSE) of all stations was greater than 0.7. The physical mechanism model and ML water level simulation models can simulate the water level in the JiaoGang Basin better. The RF and GBDT models considerably outperform the physical mechanism model in terms of the peak simulation errors and peak present time errors, and the fluctuations of the ML water level simulation models (RMSE and NSE) are minor compared to those of the physical mechanism model.