History Matching Using Hybrid Parameterisation and Optimisation Methods

Abstract

Abstract Reservoir models are commonly used in the oil and gas industry to predict reservoir behaviour and forecast production in order to make important financial decision such as reserves estimations, infill well drilling, enhanced oil recovery schemes, etc. Conditioning reservoir models to dynamic production data is known as history matching, which is usually carried out in an attempt to enhance the predicted reservoir performance. Uncertainty quantification is also an important aspect of this task, and encompasses identifying multiple history matched models, which are constrained to a geological concept. History matching and uncertainty quantification can be accomplished by identifying and using efficient and speedy optimisation techniques. The assisted history matching practice usually includes two practices; the first of which is parameterisation, which consists of reducing the number of matching parameters in order to avoid adjusting too many variables with respect to the amount of production data available. A challenging situation results from over-parameterisation, in addition to an ill-posed formulation of the inverse problem. The second process involves optimisation, which aims at solving the inverse problem by reducing a misfit or objective function that defines the difference between simulated and production data. The main challenges of optimisation are local minima solutions and premature convergence. The success of optimisation is greatly dependent on the parameterisation strategy used. These algorithms that analyse various parameterisation methods, combined and examined with diverse optimisation algorithms lead us to suggest novel hybrid approaches addressing the two processes of assisted history matching. We propose a multistage combined parameterisation and optimisation history matching technique. Hybridisation of parameterisation and optimisation algorithms when designed in an optimum manner can combine advantageous features of each method. This consisted of combining random initial parameter population by means of a wide parameter search space optimiser at early stages with initial models chosen from the best history matched models of previous stages based on the initial parameter distribution with a fine tuning optimisation algorithm at later stages. The re-parameterisation at the beginning of each stage of a hybrid algorithm assists the process in escaping local minima and prevents premature convergence. The general design of these algorithms is to initialise with simple parameterisation methods and wide spread search algorithms, in which parameterisation zoning is increased and the parameter search space is reduce with time. These hybrid algorithms allow for consistent and effective parameter search space definition in which more than one minimum can be reached, further reduce the misfit after an initial convergence has been reached, improve efficiency by accelerating the optimisation process saving valuable computing time and consequently, improved results are achieved. We also show that these hybrid algorithms can be the basis of an uncertainty range with improved predictability models when benchmarked with the Brugge synthetic model. In the case of a three stage hybrid algorithm, the misfit reduction in some cases can be improved by up to 50% relative to the first guess model, while the efficiency improvement of a hybrid algorithm with a stopping criterion saves up to eight hours for small models such as the Brugge model and an estimated 100 hours for larger models with up to 50,000 active gridcells. Finally, a recommended hybrid algorithm design for similar cases is established. We also prove that the results are independent of the first guess models used for history matching when analysed with the Brugge benchmark model. The hybrid methods offer a novel technique that incorporates effective parameterisation which defines an optimal parameter search space, and at the same time does not compromise the effectiveness of the misfit minimisation which leads to better predictive capabilities.