Deep-Learning-Based Surrogate Reservoir Model for History-Matching Optimization

Abstract

Achieving a high-quality history match is critical to understand reservoir uncertainties and perform reliable field-development planning. Classical approaches require large uncertainty studies to be conducted with reservoir-simulation models, and optimization techniques would be applied to reach a configuration where a minimum error is achieved for the history match. Such techniques are computationally heavy, because all reservoir simulations are run in both uncertainty studies and optimization processes. To reduce the computing requirements during the optimization process, we propose to create a robust deep-learning model based on the hidden relationships between the uncertainty parameters and the reservoir-simulation results that can operate as a surrogate model for computationally intensive reservoir-simulation models. In this paper, we present a workflow that combines a deep-learning, machine-learning (ML) model with an optimizer to automate the history-matching process. Initially, the reservoir simulator is run to generate an ensemble of realizations to provide a comprehensive set of data relating the history-matching uncertainty parameters and the associated reservoir-simulation results. This data is used to train a deep-learning model to predict reservoir-simulation results for all wells and relevant properties for history matching from a set of the selected history-matching uncertainty parameters. This deep-learning model is used as a proxy to replace the reservoir-simulation model and to reduce the computational overhead caused by running the reservoir simulator. The optimization solution embeds the trained ML model and aims to deliver a set of uncertainty parameters that minimizes the mismatch between simulation results and historical data. At each optimization iteration, the ML model is used to predict the well-level reservoir-simulation results. At the end of the optimization process, the optimal parameters suggested by the optimizer are then validated by running the reservoir simulator. The proposed work achieves high-quality results by leveraging advanced artificial-intelligence techniques, thus automating and significantly accelerating the history-matching process. The use of uncertainty parameters as input to the deep-learning model, and the model's ability to predict production/injection/pressure profiles for all wells is a unique methodology. Furthermore, the combination of the deep-learning surrogate reservoir model with optimization methods to resolve history-matching problems is advancing the industry's practices on the subject.