Waterflooding Optimization under Geological Uncertainties by Using Deep Reinforcement Learning Algorithms

Abstract

Abstract Recently, substantial technical progress has been made to solve complex tasks in the field of artificial intelligence (AI) by incorporating deep neural networks into reinforcement learning (RL). In this paper, four state-of-the-art deep RL algorithms are applied to optimize the net present value (NPV) of waterflooding (WF) under geological uncertainties by adjusting the water injection rate. They include the deep Q-network (DQN), double DQN (DDQN), dueling DDQN, and deep deterministic policy gradient (DDPG). A set of fifty reservoir realizations are generated by using a geostatistical technique to account for the geological uncertainties. It is found that the deep RL algorithms can optimize the WF in a 3-D 3-phase (oil-water-gas) reservoir under geological uncertainties. More specifically, both DQN and particle swarm optimization (PSO) converge to the same highest NPV, whereas the other three deep RL algorithms can find some local optimum NPVs due to the exploration-exploitation problem. DDPG converges faster than PSO and requires the least numerical simulation runs among all deep RL algorithms. The optimum water injection rate determined in the consideration of geological uncertainties not only increases the expected NPV but also reduces its standard deviation. The optimum WF starting time is found to be in the middle of the primary production. In this way, the solution-gas drive is continued and the water-cut is decreased. The production performances are compared under three different water injection scenarios: no-control, reactive-control, and optimum-control. The optimum-control scenario achieves a low water-cut and a stable oil production rate.