Application of probabilistic decline curve analysis to unconventional reservoirs

Abstract

This work presents the various probabilistic methodology for forecasting petroleum production in shale reservoirs. Two statistical methods are investigated, Bayesian and frequentist, combined with various decline curve deterministic models. A robust analysis of well-completion properties and how they affect the production forecast is carried out. Lastly, a look into the uncertainties introduced by the statistical methods and the decline curve models are investigated to discover any correlation and plays that otherwise would not be apparent. We investigated two Bayesian methods - Absolute Bayesian Computation (ABC) and GIBBS sampler - and two frequentist methods - Conventional Bootstrap (BS) and Modified Bootstrap (MBS). We combined these statistical methods with five empirical models - Arps, Duong, Power Law Model (PLE), Logistic Growth Model (LGA), and Stretched Exponential Decline Model (SEPD) - and an analytical Jacobi 2 theta model. This allowed us to make a robust comparison of all these approaches on various unconventional plays across the United States, including Permian, Marcellus, Eagle Ford, Haynesville, Barnett, and Bakken shale, to get detailed insight on how to forecast production with minimal prediction errors effectively. Analysis was carried out on a total of 1800 wells with varying production history data lengths ranging from 12 to 60 months on a 12-month increment and a total production length of 96 months. We developed a novel approach for developing and integrating informative model parameter priors into the Bayesian statistical methods. Previous work assumed a uniform distribution for model parameter priors, which was inaccurate and negatively impacted forecasting performance. Our results show the significant superior performance of the Bayesian methods, most notably at early hindcast size (12 to 24 months production history). Furthermore, we discovered that production history length was the most critical factor in production forecasting that leveled the performance of all probabilistic methods regardless of the decline curve model or statistical methodology implemented. The novelty of this work relies on the development of informative priors for the Bayesian methodologies and the robust combination of statistical methods and model combination studied on a wide variety of shale plays. In addition, the whole methodology was automated in a programming language and can be easily reproduced and used to make production forecasts accurately.