Experiences With Automated History Matching

Abstract

Matching historical performance of a reservoir is a time-consuming exercise with uncertain outcome. The current practice of manual matching is subjective and goodness of the match owes largely to the experiences of the team members involved in a study and to the quality and quantity of various input and observed data. This paper describes approaches taken to speed up the history matching effort during the course of replicating multiyear performance of two West African reservoirs. The study entails validation of production data, and matching of single-well and full-field performances with two automated-history-matching (AHM) approaches based on the Gauss-Newton algorithm. These finite-difference results obtained with AHM are contrasted with those obtained from the traditional manual approach. Results show that both AHM approaches shorten length of a study, and preserve objectivity of a history match. The use of grid coarsening results in impressive speed gain that can quickly provide clues whether adjustments of certain parameters are justified for some wells. That multiple parameter adjustments can be made within specified bounds with a few iterations in a 24-hour run is a profound improvement in our speed of learning of key reservoir flow behavior. This speed, in turn, guides the study in a path that leads to rapid conclusion in a fraction of time taken in a traditional study. Introduction History matching, by nature, is an ill-posed inverse problem. To complicate the picture, manual history matching (MHM) is a subjective, labor-intensive process. That is because MHM involves one-factor-at-a-time adjustment of static reservoir parameters. Practically speaking, the quality of a MHM exercise is driven largely by the experience of the engineers and the budget, among other intangibles. Consequently, the reliability of forecasted results is often very short-lived, and the business decisions made have large uncertainty. Yet, flow simulation remains the most potent tool to search for infill potentials and/or instituting other recovery processes in producing fields. Equally important, acquisition of real-time data overwhelms MHM and rapid turnaround for timely intervention becomes infeasible. Therefore, increasing demands are being made to develop tools that speed up the history-matching process and interject objectivity at the same time. Attempts have been made since the mid 1960's 1,2 to automate history matching. Solving the single-phase flow problem for transient-pressure tests3–5 has had good success. However, the same is not true while tackling 3D, three-phase flow problems. Difficulties arise owing to highly nonlinear nature of the problem, and large computational requirements needed to solve a sizable field problem. Although significant computational and solver efficiencies have been gained over the past four decades, ever-increasing size of geostatistical earth models have continued to challenge the computational-speed issue. Historically, several approaches have emerged to do automated matching for three-phase, full-field problems. The common thread to all of them is rooted in minimizing an objective function, which typically includes a sum of production rate mismatch terms squared. Some of these approaches include the use of steepest descent,1 Gauss-Newton for single-phase2,6–8 and multiphase9,10 flow, and optimal control theory for single-phase,6,11 two-phase,8,12,13 and three-phase14 flow problems. Successful implementation15 and use16 of the gradient-based approach has also been reported. In general, these nonlinear regression methods require evaluation of sensitivity coefficients. The sensitivity coefficients are partial derivatives of the time-dependent reservoir parameters, such as pressure and saturation, with respect to static reservoir parameters, such as permeability or azimuth of a geostatistical variogram. Numerical differentiation is a typical way of obtaining these coefficients at each iteration of regression for each estimate of a parameter. Therein lays the difficulty of gaining computational efficiency for a sizable 3D, three-phase flow problem with many parameters. One way to circumvent this issue pivots around evaluating the coefficients analytically.17

Interpreting Transient Tests in High-Permeability, Gas-Condensate Reservoirs

“…Adoption of an automated-matching technique 13,14 paved the way for rapid evaluation. We used the permeability value, obtained from the analytic approach, directly into the numeric model as an initial value.…”

Interpreting Transient Tests in High-Permeability, Gas-Condensate Reservoirs

“…For flow simulations, the gas rates at each of the three drawdowns were set at the nominal constant rates reported during the DST. We sought history matching with an automated history-matching program, 13,14 which uses an implementation of the Gauss-Newton algorithm. The calculated total skin (s t1 , s t2 , and s t3 ) were plotted against the nominal rates, in the manner shown in Fig.…”

Aquifer Matching With Material Balance Using Particle Swarm Optimization Algorithm – PSO

“…The material balance model can be used to quality check production data, estimate original hydrocarbon in place (OHIP) and identify the predominant drive mechanism of the petroleum reservoir. In order to estimate the OHIP and identify the main drive mechanism, the material balance model must be 'tuned', to a reasonable extent, to match the historical data of the reservoir -a process refered to as history matching (Kabir, Chien and Landa, 2003).…”

“…History matching is an ill-posed inverse problem and as such using a single deterministic model as a tool for accurate uncertainty quantification and production forecast might not produce an accurate result. Several approaches that result in a single optimum model has been reported.Some of these approaches include the use of steepest descent, Gaus Newton (Kabir, Chien and Landa, 2003) and Levenberg-Marquardt (Ukwu and Onyekonwu, 2014). The objective of modern history matching techniques is to generate an ensemble of solution that match the historical data of the reservoir thereby having a broad picture of the possible values of the reservoir parameter at play in the dynamics of the system.This method has the potential to capture the low and high uncertainty bandwidth, thus making it possible to estimate the confidence interval of the reservoir parameters with a high degree of accuracy.…”