A Novel Approach and Workflow for Reservoir Model Uncertainty Analysis

Abstract Reservoir model-based uncertainty analysis (UA) has been widely applied in the oil industry for decades, but it is still facing some challenges, e.g., integration and visualization of multi-dimensional (4D++) categorical and continuous input parameters. Workarounds, simplifications, or neglect have been utilized for tentative solutions. This paper presents a novel approach/workflow in the aim of attacking these challenges in reservoir model UA. Experimental design improvement for UA is not the focus of this paper. The proposed workflow was successfully applied to Kashagan, a giant carbonate oil field in Caspian Sea that consists of a matrix dominated platform and a non-matrix (NM) dominated rim. Despite 7 years of production/injection history, a wide range of uncertainties exist in reservoir characterization. To understand and mitigate their impact on field performance the subsurface uncertainties, a UA study was conducted. Categorical parameters (subsurface structure, fracture/karst distributions/connectivity, rock quality regions, reservoir heterogeneity etc.) and continuous parameters (matrix/non-matrix compressibility, dual-porosity/dual-permeability coupling parameter, aquifer size/connectivity, permeability/porosity, Kv/Kh ratio etc.) were used to characterize the subsurface uncertainties. Dynamic responses within 60-year forecast (oil/gas/water production/injection rates/ratios/cumulatives, recovery factor, oil in place, elapsed time, and objective function for field total/regions) were selected for analysis. Around 1000 simulation models were designed using standard DoE methods. Statistical analysis was conducted to identify linearity/non-linearity effects, correlations between input parameters, their statistical significances, and possible inflations during regression analysis. Multidimensional scaling was used to visualize the complex input-response relationship, QC&QA the responses and identify behaviors/characters (e.g., clusters) inside of the uncertainty space, which is not apparent in the original space. It opens a window into what the uncertainty scape is like and significantly improves the robustness of P10-P50-P90 (low, mid and high) models’ selection.