Production forecasting sits at the core of reservoir engineering workflows, required for critical tasks across the asset lifecycle. Machine learning (ML) offers distinct advantages for production forecasting, including improved accuracy and ease of automation. In this paper, we present the results of a ML-based forecasting pilot test. ML-based approach shows accuracy on par with or better than Modified Arps method for both pre-drill (future) and post-drill (existing) wells. We discuss the advantages of the ML-based approach as well as our vision for an integrated, automated production forecasting workflow.
We compared traditional Arps-based Decline Curve Analysis ("DCA") forecasts against ML-based forecasts. The latter utilized autoregressive (AR) and extra randomized trees (ETR) algorithms. This multi-variate ML forecasting workflow utilizes geology, spacing, completion variables, and historic production to forecast future production. We independently hyper-parameterized and trained models to forecast at 5- or 30-day intervals, for pre-drill and post-drill wells, and for each production stream (oil, gas, water). The production data in post-drill wells before the budget cycle date was split into training and test datasets; these robust models were then used to generate a series of forecasts aligned with budget cycles. Back-testing with production acquired after the budget cycle date provided further evidence that ML-based forecast, on average, outperforms or is on par with Modified Arps method.
On the asset level, in the pre-drill wells, the Modified Arps method shows better performance in the first sixty days, but ML-based forecasts show improved accuracy beyond that. In the post-drill wells, the ML-based models show better accuracy for oil forecasts. For gas and water, the ML forecasts are more accurate over the first 90-120 days, though less accurate at later time. The outperformance of ML-based forecast is largely attributed to its multi-variate approach, which DCA methods are incapable of. The results from this pilot provide confidence to integrate ML methods in the quarterly/annual forecasting and reserves estimation processes.
The work shows that ML-based solutions for production forecasting can match, and in certain cases exceed, the performance of simple curve fitting DCA method. ML-based methods for production forecasting, pad optimization, and budget planning are promising not only for their accuracy, but also for their speed, automation and low cost. This provides an opportunity for reservoir engineers to maximize their time on optimization workflows, which are critical in reducing costs and improving profitability of shale assets, where thousands of wells will be drilled to fully develop a field. This interdisciplinary work includes contributions from reservoir engineering, geoscience, data science, and software engineering.
