A distance is defined as a measure of dissimilarity between two reservoir models. There have been many distances proposed for fast modeling. However, some distances cause distortion or loss in original permeability distribution of models. To avoid such problems, this study proposes a pattern recognition based distance.
The distance is defined by the difference of correlation coefficients between ensemble models. From multi-dimensional scaling, initial 400 ensembles are presented on 2D plane using the distance. Then 10 groups are made by K-medoids clustering. After comparing oil production from each centroid and that of the reference field, 100 models are selected around the best centroid.
We validate the clustering by comparing the uncertainty range of 100, 50, and 20 ensemble members sampled from the initial 400 models in box plots and cumulative distribution functions. For a history matching and reservoir characterization, ensemble smoother is applied to the 100 models selected. The proposed method takes only 25% time for simulation showing reliable results compared with the initial 400 models.