Application of Distance Based Generalized Sensitivity Analysis and Bayesian Inversion Method in Simulation of a Giant Fractured Carbonate Reservoir with Unstructured Grids

Abstract

Multi-scale natural fractures bring challenges in geological modeling and flow simulation of carbonate reservoirs. History matching is extremely difficult due to significant heterogeneities and uncertainties, especially for those wells identified as "dominated by fractures". A novel, systematic approach is applied to model fractures explicitly, to perform flow simulation efficiently and eventually to match the production history accurately. First, a discrete model using unstructured triangular grids were built to fully resolve the geometry and distribution of large-scale fractures. Then the contribution of small-scale fractures was modeled using flow-based upscaling algorithms to yield enhanced porosity and permeability of matrix grid cells. Finally, the connectivity list was calculated for each pair of matrix-matrix, matrix-fracture, and fracture-fracture cells for flow simulation. Then the Distance-based Generalized Sensitivity Analysis (DGSA) method is applied to evaluate the sensitivity of the uncertain parameters in the reservoir model. Conditioning with the well production history as "given" information, the Bayesian inversion method is employed to reduce the uncertainty of fracture properties including exact position, length, and permeability etc. The entire workflow/approach was applied to a gigantic, naturally fractured reservoir with multi-billion-barrel oil reserves in Middle East. More than five hundred large-scale fractures are characterized in the simulation model explicitly using triangular prism grids. The resulted simulation model contains over 800,000 unstructured cells. It takes only one hour on a single CPU core to simulate the entire production history of over three decades for more than 100 production wells. The high simulation efficiency facilitates sensitivity analysis and history matching in which more than one thousand cases are simulated. In the meantime, due to the explicit representation of large-scale fractures, the rapid water breakthrough in some of the producers could be captured much more accurately than standard dual-porosity dual-permeability (DPDP) models. In the history matching process, the uncertainties of the sensitive parameters including most fracture and some matrix properties are systematically reduced following the Bayesian inversion method. The history-matched fracture network and matrix properties provides an accurate and efficient simulation model for future prediction and infill well optimization.