Real-time dynamic optimization method for waterflooding development based on data-driven technology and numerical simulation

Abstract

In the middle and late stages of waterflooding development, conventional manual regulation and standalone numerical simulation suffer prominent drawbacks, including delayed control and low optimization efficiency. Pure data-driven prediction models struggle to simultaneously satisfy the dual requirements of high prediction accuracy and fast online computation. This paper constructs a dual-drive optimization framework coupling an LSTM surrogate model with a fully implicit finite-difference reservoir numerical simulator to realize real-time regulation of injection-production dynamics. A nested dual-loop collaborative correction strategy and a Bayesian sliding-window online adaptive updating mechanism are proposed to effectively suppress cumulative prediction bias generated during long-term operation of purely data-driven models. Taking the cycle net present value (NPV) as the economic optimization objective, a constrained optimization model containing multiple types of reservoir engineering limits is established, and an improved particle swarm optimization (PSO) hybrid algorithm coupled with adjoint gradients is designed to solve high-dimensional well group control variables. Comparative convergence tests and field block verification demonstrate that the proposed dual-drive framework achieves a balanced tradeoff among prediction accuracy, computational cost, and long-term extrapolation robustness. This study forms a complete closed-loop intelligent regulation workflow applicable to high-water-cut reservoirs with strong heterogeneity.