Uncertainty quantification of unconventional reservoirs using assisted history matching methods

A hallmark of unconventional reservoirs is characterization uncertainty. Assisted History Matching (AHM) methods provide attractive means for uncertainty quantification (UQ), because they yield an ensemble of qualifying models instead of a single candidate. Here we integrate embedded discrete fracture model (EDFM), one of fractured-reservoirs modeling techniques, with a commercial AHM and optimization tool. We develop a new parameterization scheme that allows for altering individual properties of multiple wells or fracture groups. The reservoir is divided into three types of regions: formation matrix; EDFM fracture groups; and stimulated rock volume (SRV) around fracture groups. The method is developed in a sleek, stand-alone form and is composed of four main steps: (1) reading parameters exported by tool; (2) generating an EDFM instance; (3) running the instance on a simulator; and (4) calculating a pre-defined objective function. We present two applications. First, we test the method on a hypothetical case with synthetic production data from two wells. Using 20 history-matching parameters, we compare the performance of five AHM algorithms. Two of which are based on Bayesian approach, two are stochastic particle-swarm optimization (PSO), and one is commercial DECE algorithm. Performance is measured with metrics, such as solutions sample size, total simulation runs, marginal parameter posterior distributions, and distributions of estimated ultimate recovery (EUR). In the second application, we assess the effect of natural fractures on UQ of a single horizontal well in the middle Bakken. This is achieved by comparing four AHM scenarios with increasingly varying natural-fracture intensity. Results of the first study show that, based on pre-set acceptance criteria, DECE fails to generate any satisfying solutions. Bayesian methods are noticeably superior to PSO, although PSO is capable to generate large number of solutions. PSO tends to be focused on narrow regions of the posteriors and seems to significantly underestimate uncertainty. Bayesian Algorithm I, a method with a proxy-based acceptance/rejection sampler, ranks first in efficiency but evidently underperforms in accuracy. Results from the second study reveal that, even though varying intensity of natural fractures cam significantly alter other model parameters, that appears not to have influence on UQ (or long-term production)