Full text loading...
-
A Novel Multilevel Method For Assimilating Spatially Dense Data
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, ECMOR XVI - 16th European Conference on the Mathematics of Oil Recovery, Sep 2018, Volume 2018, p.1 - 12
Abstract
Inverted time-lapse seismic data are rich with respect to information about reservoir fluid flows. Ensemblebased data assimilation (EDA) of such spatially dense data into reservoir models requires a sufficiently large number of degrees of freedom (DOF). The DOF in straightforward EDA equals the ensemble size, E. Only a moderately sized E is, however, computationally feasible for large reservoir models. To increase the DOF, localization is routinely applied, but successful localization requires preconceived knowledge of the specific case and substantial manual effort. Alternative methods for increasing the DOF are therefore desirable. The large imbalance between data-space size (DSS) and DOF for problems with spatially dense data emphasizes this further.
We have considered generic methods for better balancing DSS and DOF. To decrease the DSS we used coarse data representation (CDR) of spatially dense data, that is, we map the data onto a regularly coarsened grid using averaging. To increase E (and, hence, the DOF) without increasing the computational cost of an ensemble forward run, we used simulations on a regularly coarsened grid with simple upscaling of reservoir properties (CGU). Results obtained with a combination of CDR and CGU, where the data and simulation grids were coarsened to the same level, were very good, but the optimal level varied from one case to another.
To avoid manual selection of an optimal level, we consider multilevel EDA using the novel Multilevel Hybrid EnKF (MlHEnKF) in combination with multilevel data representation (MDR) on a sequence of regularly coarsened grids. The resulting EDA method - MlHEnKF with MDR - can be applied in conjunction with localization, if desired. Assimilating inverted time-lapse seismic data in a reservoir-history-matching example, we assess the performance of the MlHEnKF with MDR by comparing the results to those obtained with a standard EDA approach.