Use of multi-objective algorithms in history matching of a real field

Conventionally, automated history matching methods operated on a single number (the misfit or match quality) which is obtained by summing the least square misfits for all the quantities of interest in a history match. Multi-objective history matching allows the algorithm to be guided by (usually groups of) match quality components, with the overall history match probability determined by summing the components at the end. This feature encourages diversity in the history match, as the algorithm can trade off matches in one area against good fits elsewhere, and has also been shown to speed up history matching in some cases.

This paper describes a comparison of single and multi-objective history matching of a medium-sized field in Western Siberia with nearly 100 wells and over 10 years of history. We compared the performance of both single and multi objective versions of Differential Evolution and Particle Swarm Optimisation in the history match phase, and assess the performance of all algorithms in forecast mode. We also compare the effectiveness of various different choices of history match parameters chosen based on reservoir engineering knowledge. Using these techniques, we are able to obtain a high-quality group level match with a relatively low number of simulation runs.

The paper also describes trade-offs between the parallelism of the history matching process and that of the simulator on both a 12 core windows workstation and a 240 core cluster. This is an important trade-off as neither of the extreme choices of using of all the cores by the simulator nor limiting each simulation to a single core and using each core for a separate history matching run is likely to be optimal.