Evaluation of a Bias Aware Ensemble Kalman Filter Using a Scaled Aquifer Transport Experimentby Joshua B. Kollat,
Patrick M. Reed,
Abstract: In this study, a bias-aware Ensemble Kalman Filter (EnKF) is applied to a scaled aquifer transport experiment (1-cm of the model equals 1-m at field scale and 1-day of model transport equals 1-year of field-scale transport) located at the University of Vermont. The scaled aquifer was constructed using layered porous media within a 2.54-m by 3.56-m by 2.43-m (10-ft by 14-ft by 8-ft) tank. The experimental porous media has been highly characterized and contains 105 sampling locations that are capable of providing concentration data at a high temporal resolution. The scaling of the tank experiment is applicable for advection dominated transport, which has been modeled using the three dimensional flow-and-transport groundwater models - MODFLOW 2000 and MT3DMS. A 19-day (19-years scaled) ammonia chloride tracer experiment was conducted with concentration data collected at all 105 sampling locations every 17.5-minutes (4.4-days scaled). Results from this research demonstrate how a known bias in the flow-and-transport initial conditions relative to the true experimental conditions can dramatically degrade the EnKF framework's ability to make forward transport forecasts that capture observed breakthrough time series. The bias-aware extension of the EnKF significantly improved the reliability of concentration breakthrough forecasts. In the long term, this research is being used to develop new simulation-optimization frameworks for designing groundwater observation networks. In support of this long-term objective, this study explores the influence of model bias on both the mean and covariance projections provided by EnKF. Moreover, this work explores how the increased computational demands associated with the bias-aware EnKF can be reduced using minimal ensemble sizes and management period formulations. This work advances beyond the commonly employed static Kalman filter formulations employed in prior monitoring design studies(i.e., spatiotemporal kriging) by making forecasts more robust to nonlinearities while also accounting for measurement uncertainties and dynamic, spatiotemporally correlated model structure errors as well as model parameter errors.
Subject Headings: Model analysis | Errors (statistics) | Kalman filters | Filters | Aquifers | Three-dimensional models | Forecasting | Dynamic models
Services: Buy this book/Buy this article
Return to search