Performance Assessment of Model Averaging Techniques to Reduce Structural Uncertainty of Groundwater Modeling

Abstract

Accurate estimates of groundwater modeling in arid regions have a crucial role in reaching a sustainable management of groundwater sources. However, groundwater modeling has been faced with different uncertainty sources; besides our imperfect knowledge, it is difficult to derive a proper prediction that can lead to reliable planning. This study aimed to improve the groundwater numerical simulations using different Model Averaging Techniques (MATs). For this, three numerical models, such as Finite Difference (FD), Finite Element (FE), and Meshfree (Mfree), were developed and their performance was verified in a real-world case study. Then various MATs including Simple Model Average (SMA), Weighted Average Method (WAM), Multi Model Super Ensemble (MMSE), Modified MMSE (M3SE) and Bayesian Model Averaging (BMA) were employed to improve the simulated groundwater level Fluctuations (outputs of three numerical models). The findings of this study demonstrated that the numerical model uncertainty is considerable and should not be neglected in the uncertainty analysis of groundwater modeling. In terms of RMSE, the lowest value of 0.148 m was obtained by Mfree while higher values of 1.355 m and 0.287 m are calculated for FD and FE respectively. In addition, the performance assessment of MATs showed a capacity to generate a skillful simulation compared to numerical predictions. Although the MMSE and M3SE (with RMSE values of 0.088 and 0.103 m) generated a desirable prediction in the majority of piezometers, they suffer from a main deficiency, such as the multicollinearity issue. From this perspective, it was concluded that the BMA produced a more reliable and reasonable prediction than other MATs.