Reservoir Model Maturation and Assisted History Matching Based on Production and 4D Seismic Data

Abstract

History matching in reservoir modelling is done increasingly with the help of (semi-) automated computer techniques, such as design of experiments (DoE), Ensemble Kalman Filtering or the adjoint-based techniques. Although such techniques will lead to accelerated estimation of the values of reservoir model parameters, they will not, by themselves, solve the often occurring problem of conceptual reservoir model inadequacies (missing faults, unidentified aquifers, etc.).Over the past years, we have developed a technique called "Model Maturation", where we use an adjoint-based history match to flag such inadequacies, and subsequently address those in an interdisciplinary dialogue (between reservoir engineer, geologist, petrophysicist, and geophysicist). Such a forced history match is not limited to production data, but can also include data from 4D seismic, saturation logs and RFT measurements. The "matured" reservoir model is subsequently subjected to a standard Assisted History Match (AHM) using, for example, DoE to optimize the values of the revised set of reservoir parameters.In this paper, we present a number of field cases, out of more than twenty worldwide by now, where we have successfully applied abovementioned workflow, thereby significantly improving the reservoir models in question, which in turn has a direct impact on the business decisions depending on those models.

Bayesian Style History Matching: Another Way to Under-Estimate Forecast Uncertainty?

“…If there would be much larger mismatch values for some data, this could either point to a systematic error in the data or to missing, or incorrect, features in the model (i.e., the mismatch is due to under-modeling or model error). The first requires removing erroneous data and the second requires adding new features or extra parameters to the model (i.e., the model should be "matured", Joosten, et al, 2014). If the data is deemed reliable and if it is not possible or practical to improve the model, the model error margin m has to be increased further, such that the renormalized data mismatch of each individual data point in the group becomes less than one.…”