An integrated closed-loop solution to assisted history matching and field optimization with machine learning techniques

An integrated closed-loop solution to assisted history matching and field optimization with machine learning techniques

Machine learning

Gasoline

Neural networks

Particle swarm optimization (PSO)

Regression analysis

Petroleum reservoir engineering

Genetic algorithms

Oil field development

Decision trees

Reservoir management

Learning algorithms

Trees (mathematics)

The traditional petroleum reservoir modeling workflow for reservoir characterization, history matching, and field development optimization requires a significant amount of computation during reservoir simulation and post-processing. In this work, we focus on incorporating various optimization, machine learning, and model-order reduction techniques to reduce the computational cost and build an integrated closed-loop workflow, in which the history matching and optimization procedures are carried out sequentially. The uncertainty in the predictions of petroleum reservoir behavior is addressed through multiple realizations of the petroleum reservoir system under study. The initial reservoir characterization is performed with a full physics flow simulator. Bayesian optimization (BO) is introduced for assisted history matching to find a solution reasonably fast with a small number of reservoir simulation runs. To determine the optimal field development strategy, several optimization algorithms are compared, including particle swarm optimization (PSO), genetic algorithm (GA), and a hybrid approach of these two, called genetical swarm optimization (GSO). The implementation of a proxy model for flow to replace full physics simulation effectively reduces the CPU time for the optimization tasks in the workflow. The proxy is built using machine learning algorithms applied to a set of simulation runs. In this work, proxy models built using simple multivariate regression (MVR), artificial neural network (ANN), and a decision tree-based algorithm named extreme gradient boosting (XGB) are compared. As a preprocessing step to both history matching and optimization, the parameter space for the reservoir model is reduced with KarhunenLoeve expansion (KLE) to make the size of the problem more manageable. The proposed integrated closed-loop reservoir management approach is demonstrated on a dataset representing a South Cowden reservoir in which waterflooding has been implemented. 2020 Elsevier B.V.

198

Journal of Petroleum Science and Engineering

2021

Chai, Zhi

Nwachukwu, Azor

Zagayevskiy, Yevgeniy

Amini, Shohreh

Madasu, Srinath

journalArticle

9204105

10.1016/j.petrol.2020.108204