Abstract
Recently, surrogate assisted evolutionary algorithms (SAEAs) are widely studied and applied for history matching problems due to surrogate models can accelerate convergence. However, most of the SAEAs lose the ability of parallel sampling due to the introduction of surrogate models, in which, a small number of potential solutions are selected for evaluation in each iteration. Generally, history matching involves a large number of numerical simulations, and the role of parallel computing cannot be ignored. To address this issue, this paper proposes a distributed surrogate system assisted differential evolution algorithm, termed DSS-DE. A distributed surrogate system (DSS) based on ensemble learning techniques is first developed, which builds a large number of basic learners before optimization, to effectively approximate different regions in the search space. Following that, performing multiple differential evolution (DE) optimizers with different mutation operators concurrently to sample a set of solutions to find as many as possible local or global optima of the data mismatch objective function. Moreover, based on the DSS prediction, a parallel infill strategy is designed to screen the potential promising solutions. Combined with the convolutional variational autoencoder (CVAE) based parameterization technique, a history matching workflow is developed. Empirical studies on two multimodal benchmark functions demonstrate that the proposed algorithm can obtain high-quality solutions on a limited computational budget. Furthermore, the proposed history matching workflow is validated on three synthetic waterflooding reservoir case studies with different geological characteristics. The test results show that the effectiveness of the proposed algorithm for history matching problems.
NSTL
Elsevier