We have developed a data driven workflow for forecasting future cumulative oil production in the unconventional asset development planning in the Permian basin. The workflow involves data ETL (exploration, transformation, and loading), machine learning (ML), uncertainty quantification, and result interpretation. The data integrated in the ML model are informed by subject matter experts from multiple disciplines, including geology, petrophysics, wells engineering, reservoir engineering, and operations/production engineering. The workflow consists of three steps: data preparation, ML model training, and result interpretation. The focus of paper is on discussing the effectiveness of machine learning algorithms in predicting well’s production performance, including the ML model accuracy, interpretability, and prediction uncertainty quantification. The workflow is successfully applied to a field case of multi-bench cube concept selection.
Constrained by computing power, traditional unconventional asset development relies heavily on localized reservoir simulations and fracture modeling. Lack of the ability to integrate sub-surface geological and production data is another challenge. As a result, it may yield high uncertainties in the results which pose challenges for decision making and choice of an optimal development strategy. Decades of unconventional operation have accumulated vast amount of data in all dimensions in the public databases and in the operating company’s proprietary database, which may provide us an independent insight about well production potentials by directly analyzing historical data and their trends. However, as expected, it is not straightforward to precisely predict the future oil production by simply looking at the data such as production rates, rock quality, operations, well spacing, etc. The relationship is simply nonlinear and multivariate. Machine learning algorithms, with its advantage in finding complex relationships within data, becomes an appropriate complementary approach to be used in substantiating our decision making.
A variety of efforts has been witnessed in recent literatures that has been using the data driven approaches for unconventional oil production prediction (Tadjer et al. 2021, Ning et al. 2022, Zhan et al. 2019, Gupta, et al. 2021, Luo, et al. 2018, Rahmanifard et al. 2022, Razak et al. 2021, Cui et al. 2022, Cross et al. 2022, Kormaksson et al. 2015, and so on). However, it is found that establishing a reliable workflow is not a trivial task as it bears the following aspects: 1) the collected data in the unconventional domain is usually noisy. It is not uncommon to see errors in the production allocations, low resolution geologic representations, human errors, and so on, 2) a well/well pad is a 3D subsurface system, and accurately representing the well system into a tabular data which is needed by ML is always an open question, 3) as observed, there is a big opportunity for the prediction accuracy to be further improved, 4) the interpretation of ML results needs to be conducted carefully, e.g. uncertainty quantifications, feature importance identifications, and their assumptions. We will present our ML workflow by discussing the above aspects.
