Feature selection in machine learning models of groundwater level and its application effect analysis

To improve the simulation performance of machine learning models for groundwater levels, four feature selection methods, including partial correlation analysis, Pearson correlation coefficient, maximum relevance-minimum redundancy (mRMR), and random forest (RF) methods, were employed to screen input parameters for three groundwater level machine learning models in the Mihuaishun Area. The simulation results before and after parameter feature selection were compared. The results show that different parameters require different feature selection methods. Groundwater level and its lagged values can be determined using partial correlation analysis, while artificial recharge and its lagged values, as well as precipitation and its lagged values, require a combination of mRMR and RF methods. Specifically, the mRMR method is more effective for selecting precipitation and its lagged values, whereas the RF method is better suited for screening artificial recharge and its lagged values. Feature selection significantly improved the simulation accuracy of the extreme learning machine (ELM) and RF models while enhancing the computational speed of the nonlinear autoregressive neural network with exogenous inputs (NARX) model. When applied to the three groundwater level machine learning models in the Mihuaishun Area, the parameter feature selection led to notable improvements that the ELM model showed a 63% reduction in root mean square error (RMSE), a 98% increase in the Nash-Sutcliffe efficiency coefficient (NSE), and a 45% improvement in the coefficient of determination (R²). The RF model achieved a 49% reduction in RMSE, a 6% increase in NSE, and a 2% improvement in R², while the NARX model demonstrated an 11-fold increase in computational speed.