Perpustakaan Digital ITB

Waterflood is one of methodology of oil recovery. It is categorized into secondary recovery stage after natural flow. Usually, waterflood is used as pressure maintenance and displace oil to production well. The implementation of waterflood allows a lot of variables to be optimized therefore the economical aspect can be maximized which one of them is injection rate. The injection rate in any timestep determines the NPV. There are a lot of methodologies to optimize the injection rate by doing sensitivities or using optimization algorithm like Particle Swarm Optimization (PSO). But the usage of PSO is not optimal because the decision making of the algorithm can’t be done sequentially. Every taken decision affects to current reservoir condition and future condition. The other problem of this algorithm is only done in certain condition. The solution of this special scenario can be optimal while its case is run. Meanwhile, if the scenario is changed, the solution of this algorithm is no longer optimal and it needs to re-run again. It is because the algorithm only use the objective function from simulation and ignoring the reservoir condition i.e pressure and water saturation of the field. Therefore, the new methodology is proposed by considering a sequential decision through framework Dynamic Programming (DP) and Deep Reinforcement Learning (DRL). This proposed methodology is considering dynamic properties of reservoir especially oil saturation and pressure. DRL is used to optimize the injection rate for every timestep and optimize it for around 20 years. The result shown that DRL can be more optimal then PSO and resulting the additional NPV equals to 3.75 MUSD.