Research on sensitivity analysis methods for production dynamics in heavy oil polymer flooding using proxy model

In heavy oil polymer flooding development, accurately identifying dominant control parameters and their impacts on production performance is crucial for enhancing development efficiency. To overcome the limitations of conventional numerical simulation methods characterized by time-consuming computations and low efficiency, this study employs a random search-optimized multilayer perceptron (RS-MLP) proxy model as its foundation. Through comprehensive comparison of eXtreme gradient boosting (XGBoost) and random forest (RF) algorithm performance, XGBoost was selected for feature importance analysis, utilizing information gain to quantitatively evaluate the influence of both static and dynamic parameters on oil rate, water cut, oil recovery factor, and pressure maintenance level. The study shows that nonlinear connections between parameters are well captured by the XGBoost model. While porosity largely controls fluctuations in oil output and water cut, oil saturation has the biggest effect on oil recovery factor and oil rate among static characteristics. With an average sensitivity of 0.265 across water cut, oil recovery factor, and pressure distribution, injection rate exhibits significant regulatory impacts on displacement dynamics for dynamic parameters. In comparison to traditional numerical simulation, the created RS-MLP proxy model improves computational efficiency by two orders of magnitude while achieving good prediction accuracy (R2 = 0.96) with a mean absolute error of 1.98. This study shows that, in some situations, data-driven techniques can successfully take the place of traditional numerical simulations to produce quick and precise sensitivity analysis. The method has great practical relevance for achieving intelligent reservoir management and offers a novel technical way for optimizing polymer flooding development in heavy oil reservoirs.