Surrogate-Assisted Intelligent Optimization of Graded-Viscosity Polymer Flooding with Dimensionality Reduction
Abstract
Polymer flooding is a key enhanced oil recovery technique in high-water-cut mature reservoirs, and its performance strongly depends on the appropriate design of polymer concentrations and slug sizes. However, current production optimization workflows largely rely on expert judgment and repeated high-fidelity reservoir simulations, resulting in high computational costs and time-consuming optimization processes that hinder the efficient design of polymer injection schemes. To address this issue, the graded-viscosity constraint is explicitly defined as part of the feasible region. On this basis, a graded-viscosity constraint-guided surrogate-assisted differential evolution algorithm with Sammon mapping for dimensionality reduction (GC-SADES) is developed for the economic optimization of graded-viscosity polymer flooding in high-water-cut reservoirs. The decision variables consist of the polymer concentrations and slug sizes for each injector at each injection stage, and the objective is to maximize the incremental cumulative net present value (ICNPV). A monotonic repair operator is embedded in the optimization process to ensure that all candidate solutions satisfy the prescribed monotonically decreasing polymer-concentration profile. The repaired samples are subsequently mapped into a low-dimensional space using Sammon mapping, where a Kriging surrogate model is constructed to approximate the relationship between the decision variables and the objective function. Finally, an LCB-based sequential infill strategy is employed to identify promising candidates for high-fidelity reservoir simulation, thereby reducing the number of expensive function evaluations. The results show that, at the same total polymer consumption, the graded-viscosity polymer flooding scheme achieves a recovery factor of 58.75% when the overall water cut reaches 98%, representing an increase of 5.90 percentage points over the constant-viscosity scheme. GC-SADES achieves an ICNPV of 0.83×108 RMB after 170 high-fidelity function evaluations, whereas SADE-Sammon requires 292 evaluations to reach the same level. After optimization, the ICNPV increases from 0.58×108 RMB for the initial scheme to 0.92×108 RMB, corresponding to an improvement of 58.62%. By accounting for the graded-viscosity constraint, this study provides an efficient method for optimizing polymer flooding schemes in high-water-cut reservoirs, enabling rapid and accurate optimization decisions for different types of high-water-cut reservoirs and showing broad potential for field application.