An Interpretable Recurrent Neural Network for Waterflooding Reservoir Flow Disequilibrium Analysis

Abstract

Waterflooding is one of the most common reservoir development programs, driving the oil in porous media to the production wells by injecting high-pressure water into the reservoir. In the process of oil development, identifying the underground flow distribution, so as to take measures such as water plugging and profile control for high permeability layers to prevent water channeling, is of great importance for oilfield management. However, influenced by the heterogeneous geophysical properties of porous media, there is strong uncertainty in the underground flow distribution. In this paper, we propose an interpretable recurrent neural network (IRNN) based on the material balance equation, to characterize the flow disequilibrium and to predict the production behaviors. IRNN is constructed using two interpretable modules, where the inflow module aims to compute the total inflow rate from all injectors to each producer, and the drainage module is designed to approximate the fluid change rate among the water drainage volume. On the spatial scale, IRNN takes a self-attention mechanism to focus on the important input signals and to reduce the influence of the redundant information, so as to deal with the mutual interference between the injection–production groups efficiently. On the temporal scale, IRNN employs the recurrent neural network, taking into account the impact of historical injection signals on the current production behavior. In addition, a Gaussian kernel function with boundary constraints is embedded in IRNN to quantitatively characterize the inter-well flow disequilibrium. Through the verification of two synthetic experiments, IRNN outperforms the canonical multilayer perceptron on both the history match and the forecast of productivity, and it effectively reflects the subsurface flow disequilibrium between the injectors and the producers.