Predicting field production rates for waterflooding using a machine learning-based proxy model

Abstract

Waterflooding, during which water is injected in the reservoir to increase pressure and therefore boost oil production, is extensively used as a secondary oil recovery technology. Tracking the extent and efficacy of waterflooding (i.e., fluid distributions) is a primary task of reservoir engineers and is traditionally achieved by running full reservoir models. In this work, we design and implement a proxy model using a conditional deep convolutional generative neural network (cDC-GAN), which can be used to quickly calculate the dynamic fluid distribution of a heterogeneous reservoir under waterflooding. Zero-sum game theory is the basis for the cDC-GAN, which includes a pair of generative discriminative models. The generative model tries to learn the relationship between input and output and makes the generated output as close as possible to the training data, while the discriminative model tries to distinguish the fake output and the real data used for training, such that the cDC-GAN learns the real data distribution at the end. In our cDC-GAN formulation, the reservoir properties (permeability distribution in this research) and forecast time information are treated as input, and water saturation is the desired output. The reservoir fluid production rate can be calculated based on the material balance principle. The most significant contribution of this work resides in training a cDC-GAN proxy model to accurately predict fluid saturation. A cDC-GAN has several advantages over the traditional full-model based workflow. First, the model parameters estimated from history matching help to improve reservoir characterization. Second, this proposed proxy model can predict the water and oil saturation distributions simultaneously, which can be used to calculate the water and oil flow rates. Third, this proposed proxy model can be used for waterflooding optimization and uncertainty analysis with far less computational effort than with the traditional method, which uses a reservoir simulator.