Ensemble Surrogate Modeling Based on Stratified Sampling Strategy for Reservoir Production Optimization

Reservoir production optimization aims to design rational control schemes for production and injection wells based on current reservoir conditions, with the goals of regulating the flow field and enhancing the recovery of remaining oil. It serves as a crucial approach to improving the economic performance of oilfield development. With the continuous advancement of intelligent analysis techniques and large-scale data utilization capabilities, machine learning-based surrogate modeling has been widely adopted to replace computationally expensive numerical simulators, thereby accelerating the optimization process. However, due to the dynamic evolution of reservoir conditions during development and the lack of prior knowledge, achieving high-fidelity surrogate modeling for unknown reservoirs under limited simulation budgets remains a significant challenge. To address this issue, a novel method based on stratified sampling and ensemble surrogate modeling (ESMSS) is proposed. A dual search mechanism is embedded to concurrently identify candidate development schemes that contribute most to model accuracy and optimization performance. An adaptive ensemble of surrogate models is then constructed based on predictive performance, which is further coupled with an evolutionary algorithm to establish an efficient optimization framework tailored to dynamic reservoir characteristics. The proposed method is validated on a representative three-dimensional reservoir model and compared against two mainstream optimization algorithms. Experimental results demonstrate that the proposed approach can achieve the highest economic benefit while maintaining optimization efficiency. The innovations in candidate scheme selection and adaptive surrogate integration provide a feasible pathway for intelligent optimization under complex reservoir conditions.