A Comparison of Three Stochastic Approaches for Parameter Estimation and Prediction of Steady-State Groundwater Flow: Nonlocal Moment Equations and Monte Carlo Method Coupled with Ensemble Kalman Filter and Geostatistical Stochastic Inversion.

Abstract:

We present a comparison of three stochastic approaches for estimating log hydraulic conductivity (Y) and predicting steady-state groundwater flow. Two of the approaches are based on the data assimilation technique known as ensemble Kalman filter (EnKF) and differ in the way prior statistical moment estimates (PSME) (required to build the Kalman gain matrix) are obtained. In the first approach, the Monte Carlo method is employed to compute PSME of the variables and parameters; we denote this approach by EnKFMC. In the second approach PSME are computed through the direct solution of approximate nonlocal (integrodifferential) equations that govern the spatial conditional ensemble means (statistical expectations) and covariances of hydraulic head (h) and fluxes; we denote this approach by EnKFME. The third approach consists of geostatistical stochastic inversion of the same nonlocal moment equations; we denote this approach by IME. In addition to testing the EnKFMC and EnKFME methods in the traditional manner that estimate Y over the entire grid, we propose novel corresponding algorithms that estimate Y at a few selected locations and then interpolate over all grid elements via kriging as done in the IME method. We tested these methods to estimate Y and h in steady-state groundwater flow in a synthetic two-dimensional domain with a well pumping at a constant rate, located at the center of the domain. In addition, to evaluate the performance of the estimation methods, we generated four unconditional different realizations that served as “true” fields. The results of our numerical experiments indicate that the three methods were effective in estimating h, reaching at least 80% of predictive coverage, although both EnKF were superior to the IME method. With respect to estimating Y, the three methods reached similar accuracy in terms of the mean absolute value error. Coupling the EnKF methods with kriging to estimate Y reduces to one fourth the CPU time required for data assimilation while both estimation accuracy and uncertainty do not deteriorate significantly.