Uncertainty quantification of channel reservoirs assisted by cluster analysis and deep convolutional generative adversarial networks

Abstract

Listen

Translate article

Reservoir characterization is to find reservoir properties of interest by combining available geological information. In channel reservoirs, flow responses are very sensitive depending on the characteristics of channels. Therefore, it is crucial to characterize the distribution and connectivity of channels to estimate future production performances.In this paper, we propose a method of uncertainty quantification on channel reservoir models. This novel method is based on a sequential combination of 3 machine learning methods: principal component analysis (PCA), K-means clustering, and deep convolutional generative adversarial networks (DCGAN). PCA and K-means clustering are used for classification of reservoir models and DCGAN replicates a group of new models based on the selected data from the PCA and K-means clustering.The cluster analysis by PCA and K-means clustering is to select some of realizations, which have similar production performance with the true response. After feature extraction by the two largest components from PCA, K-means clustering makes several groups among which we use silhouette method to determine the number of clusters. Then, we can find a group of models showing similar production performances with the true in an efficient way by using a representative model from each cluster.Based on the selected models, we train DCGAN to generate new realizations that do not exist among the initial models. After the cluster analysis on the regenerated models to filter out again, we can get calibrated models without using any conventional inverse algorithm. We apply the proposed method in 2 two-dimensional channel reservoir cases. In the both case, the proposed method reduces prediction uncertainty as well as characterizes channels’ pattern and connectivity reliably.