The reanalysis of a new strategy for groundwater level prediction using combined simulation of machine learning and Muskingum methods under ecological water replenishment

Abstract

Due to its multi-functionality, ecological water replenishment (EWR) has been an important measure for restoring aquifers. However, suitable prediction methods need to be selected for the unique fluctuation exhibited by groundwater level (GWL) in the process of EWR. This study employed a novel coupling model, RFR-MUS, to address this issue. RFR-MUS captured the important variable of river infiltration, which had a significant impact on GWL, and was able to achieve reliable prediction. Besides, the generalization ability, physical interpretability and application potential of RFR-MUS were explored. The results showed that the Nash-Sutcliffe efficiency coefficient (NSE) values of the predicted GWL during unknown EWR periods for more than 50% observation wells were higher than 0.6. And NSEs of the predicted GWL for 12 unknown observation wells range between −0.08 and 0.95, and the larger area affected by EWR, the higher accuracy of the simulated results. The analysis of interpretability ability of RFR-MUS is conducted by comparing the differences in explaining variable contribution with a physics-based model. The results showed that the total contribution of variables EWR and time was similar to that of EWR in physics-based model, with the highest contribution rate exceeding 95%. In addition, the results of shapley additive explanations revealed that RFR-MUS could capture the dynamics of GWL impacted by EWR, which initially increases and subsequently diminishes over time. In several verification scenarios, the response of GWL simulated by RFR-MUS to EWR was consistent with observation data, in which the maximum rise of GWL would be higher than 20 m. This study can provide a new strategy for integrating physics knowledge into ML and can also be referred as a method for GWL prediction during EWR.