A Hybrid Data-Physics Framework for Reservoir Performance Prediction with Application to H2S Production

Abstract

Summary Model-based reservoir management workflows rely on the ability to generate predictions for large numbers of model and decision scenarios. When suitable simulators or models are not available or cannot be evaluated in a sufficiently short time frame, surrogate modeling techniques can be used instead. In the first part of this paper, we describe extensions of a recently developed open-source framework for creating and training flow network surrogate models, called FlowNet. In particular, we discuss functionality to reproduce historical well rates for wells with arbitrary trajectories, multiple perforated sections, and changing well type or injection phase, as one may encounter in large and complex fields with a long history. Furthermore, we discuss strategies for the placement of additional network nodes in the presence of flow barriers. Despite their flexibility and speed, the applicability of flow network models is limited to phenomena that can be simulated with available numerical simulators. Prediction of poorly understood physics, such as reservoir souring, may require a more data-driven approach. We discuss an extension of the FlowNet framework with a machine learning (ML) proxy for the purpose of generating predictions of H2S production rates. The combined data-physics proxy is trained on historical liquid volume rates, seawater fractions, and H2S production data from a real North Sea oil and gas field, and is then used to generate predictions of H2S production. Several experiments are presented in which the data source, data type, and length of the history are varied. Results indicate that, given a sufficient number of training data, FlowNet is able to produce reliable predictions of conventional oilfield quantities. An experiment performed with the ML proxy suggests that, at least for some production wells, useful predictions of H2S production can be obtained much faster and at much lower computational cost and complexity than would be possible with high-fidelity models. Finally, we discuss some of the current limitations of the approach and options to address them.