LAUSR

Abstract Ensemble Smoother (ES) is one of popular reservoir characterization methods in petroleum engineering. It updates unknown reservoir parameters by integrating available data and utilizing multiple number of models, known as ensemble. In addition, ES is much faster than Ensemble Kalman filter (EnKF) due to one global update. However, ensemble-based methods have two key assumptions: Sufficient number of ensemble members is needed and the mean of the ensemble should be the best estimate among the realizations. Therefore, for reliable characterization using EnKF or ES, there should be enough number of well-designed initial models, reflecting true reservoir properties. The objective of this study is to reduce number of ensemble members but preserve prediction quality at the same time. We use principal component analysis (PCA) for managing high dimensional data. Total 400 ensemble members are projected on 2 dimension (2D) principal component plane and separated into 10 clusters by K-means clustering. Production histories of 10 candidate models, one from each cluster, are compared to the observed data to find out the best model for the reference. As a result, good initial ensemble models can be selected near the best model on the plane and they are used in ES. We check impacts of the sampling method in channel reservoirs by comparing 400 ensemble members with 200, 100, 50, 20, and 10 models sampled. In consideration of both time and prediction quality, around 100 are desirable numbers for uncertainty quantification and history matching. The cases with 50, 20 and 10 ensemble members may show wrong results for channel reservoirs, since EnKF and ES require more models for reliability in spite of the sampling effect. Moreover, it has been commented that ES is vulnerable to overshooting and filter divergence problem due to its global update. By selecting well-designed initial ensemble members, however, we can get better production forecasts using ES over EnKF while total simulation time is reduced about 93%. We confirm that the proposed method is effective even in the case with uncertain reservoir information and a field-scale model, PUNQ-S3. ES with the PCA-assisted model selection enables efficient history matching with a small number of ensemble realizations and one assimilation only.