An Interpretable Recurrent Neural Network for Waterflooding Reservoir Flow Disequilibrium Analysis

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