Double deep q-network no método de recuperação avançada injeção de água em um campo de petróleo

Abstract

It is necessary, for the best oil production, the constant development of new alternatives for the exploitation of the fields. The need to optimize the factors involved in this process requires great care in all the proposed recommendations. Among the elements that make up oil exploration, the following stand out: Number of wells, space between them, production/injection grid model, fluid injection system, among others. This work aims to present the development and application of an intelligent system based on the Deep Reinforcement Learning technique in oil reservoirs submitted to the advanced waterflooding recovery method. The simulation was carried out with the mathematical simulator STARS (Steam Thermal and Advanced Process Reservoir Simulator) from the CGM (Computer Modeling Group) group, considering a homogeneous semi-synthetic reservoir with characteristics similar to those found in Northeast Brazil. The applied algorithm was the Double Deep Q-Network (DDQN), which consists of an association between a deep learning network and the Q-learning algorithm and aims to find favorable operating conditions, aiming to maximize the Net Present Value (NPV) and the significant increase in the Recovery Factor, with actions to increase or not the water injection flow rate at the beginning of production within a production horizon estimated at 240 months (20 years). The use of the algorithm provided the optimal operating conditions that enabled significant increases in the field’s recovery factor, as well as in the NPV and, consequently, the profitability, with a drop in costs with water injection, treatment and disposal of produced water, thus generating an increase in the project’s viability time.