Reservoir characterisation using artificial bee colony optimisation

Abstract

To obtain an accurate estimation of reservoir performance, the reservoir should be properly characterised. One of the main stages of reservoir characterisation is the calibration of rock property distributions with flow performance observation, which is known as history matching. The history matching procedure consists of three distinct steps: parameterisation, regularisation and optimisation. In this study, a Bayesian framework and a pilot-point approach for regularisation and parameterisation are used. The major focus of this paper is optimisation, which plays a crucial role in the reliability and quality of history matching.Several optimisation methods have been studied for history matching, including genetic algorithm (GA), ant colony, particle swarm (PS), Gauss-Newton, Levenberg-Marquardt and Limited-memory, Broyden-Fletcher-Goldfarb-Shanno. One of the most recent optimisation algorithms used in different fields is artificial bee colony (ABC). In this study, the application of ABC in history matching is investigated for the first time. ABC is derived from the intelligent foraging behaviour of honey bees. A colony of honey bees is comprised of employed bees, onlookers and scouts. Employed bees look for food sources based on their knowledge, onlookers make decisions for foraging using employed bees' observations, and scouts search for food randomly.To investigate the application of ABC in history matching, its results for two different synthetic cases are compared with the outcomes of three different optimisation methods: realvalued GA, simulated annealing (SA), and pre-conditioned steepest descent. In the first case, history matching using ABC afforded a better result than GA and SA. ABC reached a lower fitness value in a reasonable number of evaluations, which indicates the performance and execution-time capability of the method. ABC did not appear as efficient as PSD in the first case. In the second case, SA and PDS did not perform acceptably. GA achieved a better result in comparison to SA and PSD, but its results were not as superior as ABC's.ABC is not concerned with the shape of the landscape, that is, whether it is smooth or rugged. Since there is no precise information about the landscape shape of the history matching function, it can be concluded that by using ABC, there is a high chance of providing high-quality history matching and reservoir characterisation.