A Hybrid Numerical Method Incorporating Machine Learning into Groundwater Level Model for Improving Simulation Accuracy

Abstract

Groundwater level (GWL) serves as a key indicator for assessing groundwater resources. Traditional numerical models for simulating GWLs typically rely on physics-based approaches, which require detailed descriptions of the hydrogeological structure and the related parameters. The data-driven models, while offering relatively high simulation accuracy, often lack interpretability. This paper proposes a hybrid numerical method framework that integrates a numerical model and a data-driven model to enhance GWL simulation accuracy. The framework is applied to the upper and middle parts of the Chaobai River alluvial fan in Beijing, China, a significant emergency water resource region. This framework is also used to predict landfill infiltration risks. The results indicate that the hybrid model significantly improves the simulation accuracy on the training set, achieving a maximum reduction of 83% in RMSE, 87% in MAE, and a maximum improvement of 56% in R², compared to the numerical model alone. During the validation period, 8 of the 11 wells demonstrated an RMSE below 1 m and R² exceeding 0.85. The prediction results indicate that from 2022 to 2028, even in scenarios with reduced withdrawal or increased precipitation, GWLs will still be below the depth of landfills, which poses no potential risk of groundwater contamination.