Fast and Reliable History Matching of Channel Reservoirs Using Initial Models Selected by Streamline and Deep Learning

Abstract

Ensemble-based methods involve using multiple models for model calibration to correct initial models based on observed data. The assimilated ensemble models allow probabilistic analysis of future production behaviors. It is crucial to use good initial models to obtain reliable history matching and prediction of both oil and water productions especially for channel reservoirs having high uncertainty and heterogeneity. In this study, we propose a fast and reliable history matching method by selecting good initial models using streamline and deep learning. The proposed method is applied to two cases of 3D channel reservoir generated by sgems and generative adversarial network (GAN). The proposed method offers predictions with accuracy improvement more than 20% for oil and 10% for water productions compared with two other model selection methods. It also reduces the overall simulation time by 75% compared to the method of using all initial models.