A comparative evaluation of global search algorithms in black box optimization of oil production: A case study on Brugge field

作者：Foroud Toomaj; Baradaran Ali; Seifi Abbas^*

来源：Journal of Petroleum Science and Engineering, 2018, 167: 131-151.

DOI：10.1016/j.petrol.2018.03.028

• Oil production optimization
• Black box simulator
• Brugge field
• Simultaneous perturbation stochastic approximation (SPSA)
• Ensemble base optimization (En-Opt)
• Particle swarm optimization (PSO)
• Pattern search (PS)
• Guided pattern search (GPS)
• Covariance matrix adaptation evolutionary strategy (CMAES)
• Differential evolution (DE)
• Self adaptive differential evolution (SADE)

We evaluate the application of eight different global search algorithms to the optimization of oil production from a mature field. Our focus is on algorithms that treat the reservoir simulator as a black box, which is the case for most commercial hydrocarbon reservoir simulators. The selected optimization algorithms have been divided in two categories. The first category consists of those algorithm that use approximated gradients, namely, simultaneous perturbation stochastic approximation (SPSA) and ensemble base optimization (En-Opt) methods. The second group includes derivative-free algorithms including particle swarm optimization (PSO), pattern search (PS), guided pattern search (GPS), covariance matrix adaptation evolutionary strategy (CMAES), differential evolution (DE) and self-adaptive differential evolution ( SADE). GPS algorithm has been recently introduced and applied in oil production optimization by the authors (Foroud and Seifi, 2016) while the other algorithms have been developed and coded in MATLAB software according to the most renowned studies in the literature.
The selected algorithms have been applied to optimization of oil production in Brugge field. This problem is a bounded NPV optimization with 640 decision variables consist of injection and production rates over 10 years of operation and 200 linear inequality constraints. The results here show that algorithms that use approximated gradients (SPSA and En-Opt) and take advantage of physical properties of the underlying problem (GPS) are superior. Algorithms with self adaptation ability such as SADE and CMAES are the second best performers on this application. In fact, SADE which is the self adapted version of DE could achieve 7.5% more NPV than ordinary DE algorithm. Finally, in this study, GPS has been overall the most efficient algorithm with lowest number of function evaluations and the second highest NPV compared to other algorithms.

• 出版日期2018-8