Traducción del resúmen del autor: Over the last decade the ensemble Kalman filter (EnKF) has attracted attention as a promising method for solving the reservoir history matching problem: updating model parameters so that the model output matches the measured production data. The method possesses unique qualities, such as; it provides real-time updates and uncertainty quantification of the estimate, it can estimate any physical property at hand and it is easy to implement. The method does, however, have its limitations; in particular, it is derived based on an assumption of a Gaussian distribution of variables and measurement errors. Several refinements have been proposed to overcome the shortcomings of the EnKF. These refinements are, however, mainly tested on synthetic cases addressing one shortcoming at a time, not containing or combining the complexity and the high nonlinearity of a real field case. In this paper, we investigate some of the refined methods on a nonlinear reservoir, the 3D, three-phase, PUNQ-S3 model. We compare the performance of the original EnKF with the performance of the ensemble square root filter (EnSRF), an EnKF method with localization, which is named the hierarchical ensemble Kalman filter (HEnKF) and the newly proposed Adaptive Gaussian Mixture filter (AGM). To the best of our knowledge, this is the first time the EnKF and the EnSRF have been compared on a high-dimensional nonlinear field case. Overall, we see that the AGM and HEnKF work better than the EnSRF and EnKF. The EnSRF seems to have a slightly better performance than the EnKF. However, the introduction of a localization procedure (as in the HEnKF) seems to be much more influential than replacing the EnKF with the EnSRF. Comparing the top two methods, the AGM is preferable over the HEnKF, both when it comes to preserving the initial geology of the ensemble and to the consistency of the predictions.

5