Summary Production-data analysis is a practice fraught with inconsistencies. In the application of any single model, the quantity of answers arrived at by experienced evaluators is often equal to the number of evaluators analyzing the data. The cause of such inconsistency is bias on the part of evaluators. Although the colloquial use of bias typically implies systematic error, in this paper, we define bias as an expression of belief by the evaluator. With the lack of recognition of bias, no means exists with which to gauge its accuracy. A method that requires explicit expression of one's bias in time/rate decline behavior can provide an objective means with which to evaluate it. In this work, we present a machine-learning method to forecast production in unconventional, liquid-rich shale and gas-shale wells. Methods were developed for probabilistic decline-curve analysis with Markov-chain Monte Carlo simulation (MCMC) as a means to quantify reserves uncertainty, to incorporate prior information (i.e., bias), and to do so quickly. We extend the existing approaches by (a) a modified likelihood-distribution function to improve “learning” of production data, (b) integration of the transient hyperbolic model (THM) to explicitly define the various flow regimes present in unconventional wells, (c) a method for construction of discretized “percentile neighborhood” forecasts, and (d) construction of type wells from an analyzed well population. The accuracy and calibration of the method are demonstrated by an analysis of 136 wells in the Elm Coulee Field of the Bakken. Quantification of change in time/rate behavior caused by completion design, and the inference of physical behavior and properties, is demonstrated with a tight oil play in the Cleveland sand formation of the Anadarko Basin, as well as a shale play in the Wolfcamp formation of the Permian Basin. We show that this implementation of supervised machine learning, in combination with well-calibrated bias, improves the estimation of uncertainty of the posterior distribution of forecasts. In addition, hindcasts performed at various time intervals result in accurate estimation of mean five-year cumulative production. We observe that the “percentile neighborhood” forecasts are reasonable fits of production data comparable to those that may be created by a human evaluator, and that the type well computed is representative of the decline behavior of the well population upon which it is based. We conclude that, given the speed and accuracy of the process, machine learning is a reliable technology as defined by the US Securities and Exchange Commission (SEC), and can significantly improve the process of production forecasting by human evaluators for most unconventional wells with consistent trends of production history.

This paper presents the use of machine learning via a multiple linear regression and a neural network to solve the complex problem of optimizing completions and well designs in the Duvernay shale. Solutions were revealed that could save over a million dollars per well, along with the potential for more than 50% improvement in well performance. This was accomplished through a workflow that rigorously analyzed the relationships between a multitude of well completion variables, generated predictions of future results, and performed optimizations for ideal outcomes. Most importantly, this workflow is not Duvernay specific, and can easily be applied to other basins and formations. This is a fundamental problem in many industries, in that a responding variable is controlled not just by one predictor variable, but by a number of predictor variables. Inferring the relationship between the responding variable and the predictor variables is then of key importance. Interactions between predictor variables, as well as noise in the data, complicate matters further. This problem can be solved with a multiple linear regression or a neural network, both of which utilizes all predictor variables together. However, care must be taken to obtain a model that is truly predictive and not a result of overfitting the data. The workflow was applied to 262 Duvernay wells, ranging from dry gas to volatile oil. No wells were excluded for operational or geological reasons, a strength of this methodology. By not excluding any wells, the model could maximize learnings and establish statistical reasons for the variances in well performance observed. The final model achieved very high predictive power, correctly predicting 78% of the variance in well performance on 52 wells the model hadn't been trained on. Conclusions were quite significant, including: Indicating virtually no benefit from more expensive fracturing procedures, such as using ceramic or resin coasted proppant, or having hybrid fluid systems, offering savings of over a million dollars per well in the Duvernay.No benefit from placing wells on an azimuth (parallel to the minimum horizontal stress) vs. a North-South orientation (~45° off azimuth). This allows potentially large savings on a land ownership system not aligned to this direction, by allowing simpler pad design in achieving the same aerial coverage of reservoir depletion.Confirming total fracture tonnage as a key driver of well performance.Suggesting fracture pump rate is associated with better well performance and should be investigated further. These conclusions would have been very difficult to derive without expensive strategic testing on numerous wells with rigorous control of the completions and geological inputs. When compared to recent well performance of six operators, the neural network predicted substantial ability to improve well performance by varying parameters under operator control. Potential improvement ranged from 19% to 97%, showing large potential improvement for all operators.

Production Data Analysis is a practice wrought with inconsistencies. In the application of any single model, the quantity of answers arrived at by experienced evaluators is often equal to the number of evaluators analyzing the data. The cause of such inconsistency is bias on the part of evaluators. While the colloquial use of bias typically implies systematic error, in this paper we define bias as an expression of belief by the evaluator. With the lack of recognition of bias, no manner exists in which to gauge its accuracy. A method that requires explicit expression of one's bias in rate-time decline behavior can provide an objective manner in which to evaluate it. In this work, we present a machine learning method to forecast production in unconventional, liquids-rich shale and gas shale wells. Gong et al [2011] developed a method for probabilistic decline curve analysis using Markov-chain Monte Carlo simulation (MCMC) as a means to quantify reserves uncertainty, incorporate prior information (i.e. bias), and to do so quickly. However, their approach resulted in limited use of discrete P10, P50, & P90 production forecasts, as these often did not align with production data. We extend their approach by a) utilizing the Transient Hyperbolic Model (THM) to represent the various flow regimes present in unconventional wells, b) a methods for construction of representative "percentile neighborhood" forecasts, c) construction of Type Curves from an analyzed well set, and d) a modified likelihood algorithm to improve the accuracy of discrete forecasts. The accuracy and calibration of the method is demonstrated by an analysis of 136 wells in the Elm Coulee Field of the Bakken. Quantification of change in rate-time behavior due to completion design, and the inference physical behavior and properties, is demonstrated using a tight oil play in the Cleveland sand formation of the Anadarko Basin, and a shale play in the Wolfcamp formation of the Permian Basin. We show that this implementation of supervised machine learning, in combination with well-calibrated bias, improves the estimation of uncertainty of the distribution of forecasts. Additionally, hindcasts performed at various time intervals results in accurate Mean 5 year cumulative production. We observe that the "percentile neighborhood" forecasts are reasonable fits of production data comparable to those that may be created by a human evaluator, and that the type curve computed is representative of the decline behavior of the wells upon which it is based. We conclude that, given the speed and accuracy of the process, machine learning is a reliable technology as defined by the SEC, and can replace the process of manual production forecasting by human evaluators for most unconventional wells with consistent trends of production history.

No abstract

We extend the numerically-assisted RTA workflow proposed by Bowie and Ewert (2020) to (a) all fluid systems and (b) finite conductivity fractures. The simple, fully-penetrating planar fracture model proposed is a useful numerical symmetry element model that provides the basis for the work presented in this paper. Results are given for simulated and field data. The linear flow parameter (LFP) is modified to include porosity (LFPꞌ=LFP√φ). The original (surface) oil in place (OOIP) is generalized to represent both reservoir oil and reservoir gas condensate systems, using a consistent initial total formation volume factor definition (Bti) representing the ratio of a reservoir HCPV containing surface oil in a reservoir oil phase, a reservoir gas phase, or both phases. With known (a) well geometry, (b) fluid initialization (PVT and water saturation), (c) relative permeability relations, and (d) bottomhole pressure (BHP) time variation (above and below saturation pressure), three fundamental relationships exist in terms of LFPꞌ and OOIP. Numerical reservoir simulation is used to define these relationships, providing the foundation for numerical RTA, namely that wells: (1) with the same value of LFPꞌ, the gas, oil and water surface rates will be identical during infinite-acting (IA) behavior; (2) with the same ratio LFPꞌ/OOIP, producing GOR and water cut behavior will be identical for all times, IA and boundary dominated (BD); and (3) with the same values of LFPꞌ and OOIP, rate performance of gas, oil, and water be identical for all times, IA and BD. These observations lead to an efficient, semi-automated process to perform rigorous RTA, assisted by a symmetry element numerical model. The numerical RTA workflow proposed by Bowie and Ewert solves the inherent problems associated with complex superposition and multiphase flow effects involving time and spatial changes in pressure, compositions and PVT properties, saturations, and complex phase mobilities. The numerical RTA workflow decouples multiphase flow data (PVT, initial saturations and relative permeabilities) from well geometry and petrophysical properties (L, xf, h, nf, φ, k), providing a rigorous yet efficient and semi-automated approach to define production performance for many wells. Contributions include a technical framework to perform numerical RTA for unconventional wells, irrespective of fluid type. A suite of key diagnostic plots associated with the workflow is provided, with synthetic and field examples used to illustrate the application of numerical simulation to perform rigorous RTA. Semi-analytical models, time, and spatial superposition (convolution), pseudopressure and pseudotime transforms are not required.