Integrated reservoir characterization using nonparametric regression and multiscale Markov random fields

This dissertation presents effective and novel techniques for permeability predictions in uncored wells from well logs using classification and regression, spatial modeling using nonparametric regression, and integrating multiresolution data from a variety of sources into fine scale reservoir models using multiscale Markov Random Field (MRF) for accurate performance forecasting.; First, a two-stage approach for permeability predictions from well logs is presented. The main idea of the proposed technique is that by pre-classifying the well log responses into several distinct clusters, or electrofacies (EF) and finding the optimum permeability correlation model for each class using non-parametric regression, more accurate permeability predictions can be obtained. We have used a conventional “unsupervised” pattern recognition technique, model-based clustering, to identify the clusters from well log responses. In predicting permeability at uncored wells, a conventional “supervised” pattern recognition technique, or discriminant analysis was used to find the EF group to which the set of well logs responses at each level is assigned. The field application shows that the proposed technique can improve permeability estimation under highly heterogeneous environments.; Second, a nonparametric approach for spatial modeling is presented for integrating well data into reservoir descriptions. Synthetic examples demonstrate the power of Locally Weighted Polynomial Regression (LWPR) as the local trend analysis and field examples illustrate the practical applicability of LWPR for integrating well data into 3-D reservoir model. For stochastic simulations, a hybrid method based on a combination of LWPR and conventional geostatistics is proposed. This method utilizes the advantages of conventional geostatistics such as reproducing the sample data and quantifying uncertainty of the estimates and exploits the power of LWPR for effectively capturing local trend in the spatial data.; Third, a hierarchical approach to spatial modeling based on MRF and multi-resolution algorithms in image analysis is presented for multi-scale data integration. This method is computationally efficient and well-suited to reconstruct fine scale non-Gaussian spatial fields from coarser, multi-scale samples and sparse fine scale conditioning data. It is easy to implement non-linear interactions between different scales. Synthetic examples demonstrate the advantages and superiority of the proposed method over conventional geostatistical techniques.