Li Yang, Wang Sheng, Li Menghan, Zhang Zhao
Geological parameters in petroleum reservoirs are highly heterogeneous, but the exploration approaches are very limited. In consequnce, geological models are usually highly uncertain. No matter conventional history matching or data space inversion, it is necessary to predict future reservoir dynamics under the uncertainty of geological models. This is conventionally done by reservoir simulation. Yet, for the large ensemble of realisattions to reflect the uncertainty of geological models, reservoir simulation is computaionally expensive. For this problem, we extend dynamic mode decomposition (DMD) to the parameter space, and conduct prediction based on a series of geological realisations reflecting the uncertainty of formation parameters. The change of reservoir flow variables tend to become gradual over time. Therefore, we build training and testing data sets, and determine the time after which the dynamic data approximately satisfies local linearity. Then DMD can be used to conduct prediction as a surrogate for reservoir simulation, in order to enhance the efficiency of prediction. The new method is validated using single and two-phase transient Darcy flow test cases.
