Quantification and reduction of groundwater model structural uncertainty based on Gaussian process regression

Nowadays, the main analysis method for groundwater model structural uncertainty is the Bayesian model averaging (BMA) method. But BMA suffers from the difficulty of model weight estimation, which makes its application infeasible. More attention is recently paid to the data-driven based model structural error analysis method. In this paper, the groundwater model structural error is statistically learnt based on Gaussian process regression, and then the parameters of the groundwater model and statistical model are identified simultaneously by combining the DREAMzs and Gaussian process regression algorithms. With this method, the uncertainty of groundwater model parameters and prediction results are analyzed. In addition, a synthetic numerical simulation of seawater intrusion in a karst fissure area and a solute transport column experiment are taken as case studies. In contrast with the uncertainty analysis without considering the model structural error, the impact of parameter compensation is significantly reduced by considering the model structural error. Moreover, the model prediction performance is also improved. Therefore, based on the model structural uncertainty analysis method proposed in this paper, the uncertainty of groundwater modeling can be reduced to some extent, as well the reliability of groundwater model prediction can be improved.