A Machine Learning Application for Field Planning

Abstract

Recently, machine learning methods have enjoyed resurgence in the oil and gas industry and provide applications to a wide variety of problems (Noshi et al. 2018). However, few address the important problem of field planning, which is the area of focus of this paper. This paper introduces a machine learning-based framework for field planning and specifically addresses the problem of well location planning. Unsupervised learning is used to understand characteristics of the data, followed by the creation of a regression model trained on available data in a marginal oilfield to develop a prediction tool for well productivity. This data-based prediction tool is used in a workflow to optimize well locations under uncertainty, and then in another workflow, using an adaptation from modern portfolio theory (MPT) (Markowitz 1952), to evaluate recommended well locations. The latter workflow enables the operator to select a portfolio of wells to maximize returns at a given risk tolerance. Both workflows are applicable to plays where drainage areas for wells are small; consequently, a well can be assumed to produce independently of other wells. This assumption is true for both early-stage shale plays and conventional fields in tight rocks. As compared to traditional reservoirs, a very large number of wells are drilled in shale plays, generating a very large amount of data that existing productivity prediction tools, such as reservoir simulators, find difficult to consume and use in practical timeframes. The presented framework is attractive for shales because it can easily handle large datasets. The framework accounts explicitly for risk resulting from uncertainty in well productivity and uses data typically acquired in the field, but not used currently for field planning.