Accurate groundwater level (GWL) prediction is crucial in groundwater resource management. Currently, it relies mainly on physics-based models for prediction and quantitative analysis. However, physics-based models used for prediction often have errors in structure, parameters, and data, resulting in inaccurate GWL predictions. In this study, machine learning algorithms were used to correct the prediction errors of physics-based models. First, a MODFLOW groundwater flow model was created for the Hutuo River alluvial fan in the North China Plain. Then, using the observed GWLs from 10 monitoring wells located in the upper, middle, and lower parts of the alluvial fan as the test standard, three algorithms—random forest (RF), extreme gradient boosting (XGBoost), and long short-term memory (LSTM)—were compared for their abilities to correct MODFLOW’s predicted GWLs of these 10 wells under two sets of feature variables. The results show that the RF and XGBoost algorithms are not suitable for correcting predicted GWLs that exhibit continuous rising or falling trends, but the LSTM algorithm has the ability to correct them. During the prediction period, the LSTM2 model, which incorporates additional source–sink feature variables based on MODFLOW’s predicted GWLs, can improve the Pearson correlation coefficient (PR) for 80% of wells, with a maximum increase of 1.26 and a minimum increase of 0.02, and can reduce the root mean square error (RMSE) for 100% of the wells with a maximum decrease of 1.59 m and a minimum decrease of 0.17 m. And it also outperforms the MODFLOW model in capturing the long-term trends and short-term seasonal fluctuations of GWLs. However, the correction effect of the LSTM1 model (using only MODFLOW’s predicted GWLs as a feature variable) is inferior to that of the LSTM2 model, indicating that multiple feature variables are superior to a single feature variable. Temporally and spatially, the greater the prediction error of the MODFLOW model, the larger the correction magnitude of the LSTM2 model.

The hydrochemistry and quality of local shallow groundwater was assessed within the Xinzhou basin in Shanxi, North China. Piper diagrams, correlation analysis, principal component analysis, chloro-alkaline indices, ion proportion diagrams, and Gibbs diagrams were used to reveal the hydrochemical characteristics and evolution mechanisms of groundwater. Besides, two indicators of sodium adsorption ratio and soluble sodium percentage, USSL and Wilcox diagrams, and water quality index models were used to evaluate the groundwater quality for irrigation and drinking. In general, groundwater in most areas of the basin is fresh water with total dissolved solid lower than 1000 mg/L. But there are salt water in some parts of the southern basin, with total dissolved solid higher than 1000 mg/L, due to industrial and domestic pollution. The hydrochemical facies of groundwater in most areas are HCO₃-Ca and mixed HCO₃-Ca·Mg·Na types, while it is HCO₃·SO₄·Cl-Na type in a small part of the basin. Nitrate pollution widely occurred in the basin because of the use of nitrogenous fertilizers. The dominant mechanism controlling the evolution of groundwater chemistry compositions was the weathering of rock minerals. It mainly reflected in the dissolution of carbonate minerals. And the carbonate dissolution is the major origins of HCO₃⁻, Ca²⁺, and Mg²⁺ in groundwater. However, Na-Ca exchange is the important source of Na⁺. Shallow groundwater was suitable for irrigation and drinking except for some southern parts of the basin. These results will be helpful for the protection and efficient management of groundwater in the Xinzhou basin.

Accurate prediction of groundwater levels (GWLs) is crucial for groundwater resource management. Traditional numerical models, such as MODFLOW, often fail to provide precise predictions due to data limitations, while machine learning models like long short-term memory (LSTM) lack interpretability. This study proposes an integrated approach that combines MODFLOW predictions with LSTM to enhance both prediction accuracy and interpretability. The performance of the coupling model (MOD-LSTM) was tested using data from ten monitoring wells located at the top, middle, and bottom of a typical alluvial fan in the North China Plain. The results demonstrate a significant improvement in GWL prediction accuracy. During the training and validation periods, 80% of the wells achieved a Nash-Sutcliffe Efficiency (NSE) greater than 0.75. During the prediction period, the MOD-LSTM model outperformed MODFLOW, with an average increase of 0.423 in the Pearson correlation coefficient (PR) and a decrease of 0.618 m in the root mean square error (RMSE); Compared to pure LSTM model, the MOD-LSTM model showed a 0.145 increase in PR and a 0.280 m decrease in RMSE. The model also effectively captured long-term trends of GWLs, seasonal fluctuations, and rising trends following river ecological water replenishment events. Furthermore, the predictions of the MOD-LSTM model can be interpreted using the MODFLOW model, due to the similar GWL dynamics of each well. The coupling model combines physics-driven modeling with machine learning, providing a novel approach for more accurate groundwater level prediction.