Drainage Strategy Optimization Using Machine Learning Methods

Abstract

ABSTRACT Oil production optimization is a crucial issue in the oil industry. Simulating different production scenarios effectively and quickly enables companies to automate and optimize production systems. This paper presents a study on developing intelligent agents to aid reservoir engineers in optimizing oil production. We propose a machine learning model for optimizing oil production over time by adjusting the pressure in oil reservoirs. The proposed model architecture uses three encoders (field values encoder, well values encoder, and 3D grid values encoder) to process input data. Using the encoder outputs, a dense neural network generates a policy function that determines how much pressure adjustment is required for each well in the oil field based on the probability distribution. We evaluate the proposed approach through experimentation. It is worthwhile to mention that, in our experiments, we had to discretize the reservoir well pressure adjustments to be able to compute them. Nevertheless, the results of the experiments show that our proposed model can learn how to optimize the reservoir well pressure with an Elo rating of 349.40 points after training over eleven generations. Also, the results show that the optimization process increases oil production by 1074.5% on a simulated test reservoir with two producers and one injector well, respectively. Although our experimental results reflect only the case of a simulated reservoir environment, we can see that our implementation has huge potential in a real oil reservoir field.