Surrogate-reformulation-assisted multitasking knowledge transfer for production optimization

Abstract

Data-driven surrogate models, which are trained by samples to replace time-consuming numerical simulations, have been widely used to solve production optimization problems in recent years. It is a challenging and meaningful subject to research advanced surrogate-model-based methods that can obtain superior optimization performance within a limited time budget. The key is to enhance the quality of each training sample, i.e., the contribution of each sample to the overall optimization performance improvement, because the acquisition of each sample requires a numerical simulation that generally costs tens of minutes or even several hours. To obtain samples with enhanced quality, a novel approach named surrogate-reformulation-assisted multitasking knowledge transfer (SRAMT) for production optimization is proposed in this research. Multiple surrogate models, which can imitate the landscape of the initial production optimization problem, are constructed with diverse samples as reformulations of the target problem. These models reflect different landscape information of the same problem and thus can be regarded as multiple associated optimization instances, which just provide a solid foundation for the subsequent process. Then, an advanced optimization method, namely multitasking optimization (MTO), is leveraged to find optimal solutions for these surrogates. MTO can handle several optimization instances simultaneously and boost the performance of each one by transferring useful knowledge among tasks. Besides, in the absence of prior knowledge about the target production optimization task, as in most situations, an approach is proposed to determine the frequency of knowledge transfer adaptively based on the similarity between surrogates to improve efficiency and stability. To verify the effect, four 100-dimensional benchmark functions and two reservoir models are tested on the method and the results are in comparison with those obtained by differential evolution (DE) algorithm and three other surrogate-model-based methods. The results show that the proposed method can achieve optimal well controls which can get the highest net present value (NPV) for target production optimization problems and have superior convergence speed. • A method based on surrogate reformulation and multitasking enhancement is proposed. • Knowledge is transferred among two data-driven surrogate models for better samples. • The transfer frequency is controlled adaptively to avoid the negative transfer. • The method is applied to four benchmark functions and two reservoir examples.