Geology-guided Quantification of Production-Forecast Uncertainty in Dynamic Model Inversion

Abstract

Abstract The presence of a large number of geologic uncertainties and limited well data typically increase the challenges associated with hydrocarbon recovery forecasting. Although recent advances in geologic modeling enable the automation of the model generation process by means of next-generation geostatistical tools, the computation of the reservoir dynamic response with full-physics reservoir simulation remains a computationally expensive task, which in practice requires considering only a few (but which?) of the many probable realizations. This paper presents a workflow that demonstrates the potential of capturing the inherent model uncertainty more accurately and assists in production-forecast business decisions. This workflow uses a history matching approach that directly interfaces the Earth modeling software with a forward simulator. It relies on the rapid characterization of the main features of the geologic uncertainty space, represented by an ensemble of sufficiently diverse history matched model realizations at the high-resolution geological scale. This workflow generates a more accurate result by obeying known geostatistics (variograms) and well constraints. We implement a multi-step, Bayesian Markov chain Monte Carlo inversion in which the proxy model is guided by streamline-based sensitivities. This process eliminates the need to run a forward simulation for each model realization, which significantly reduces the computation time. Efficient model parameterization and updates in the wave-number domain, based on discrete cosine transform (DCT), is used for fast characterization of the main features of the geologic uncertainty space, including structural framework, stratigraphic layering, facies distribution, and petrophysical properties. The application of the history matching workflow is demonstrated with a case study the combines the geological model with approximately 900K cells, four different depositional environments, and 30 wells with a 10-year waterflood history. Finally, the method is described to dynamically rank the reconciled model realizations to identify the highest potential of capturing bypassed oil and to optimize business decisions for implementing improved oil recovery (IOR). The main features include the following: Calculation of pattern-dissimilarity distances, which distinguish two individual model realization in terms of recovery response Deployment of very fast streamlined simulations to evaluate distances Application of pattern-recognition techniques to assign several realizations, representative for production forecasting, to full-physics simulation Derivation of the probability distribution of dynamic model responses (e.g., recovery factors) from the intelligently selected simulation runs