Diagnosing reservoir model deficiency for model improvement

Abstract

Models must be checked for consistency with actual observations before they can be used for forecasting. This can be done either before data are assimilated through calibration, or after assimilation of data. In this paper we describe a process for model checking before calibration that goes beyond rejection or acceptance of the model. Simply knowing that a model is inconsistent with the truth is not sufficient. One needs information to help improve the model so that it is not demonstrably inconsistent. This can sometimes be accomplished by identifying the source of the model discrepancy in data space, i.e., which data make the largest contribution to model discrepancy. Once these data have been identified, one can attempt to attribute that discrepancy to a deficiency of the model: missing parameters, missing physical processes, mischaracterization of uncertainty. In order to identify deficiencies that are most likely to lead to model improvement, we focus on a methodology that identifies pairs and triplets of data that result in rejection of the model. The approach is illustrated on the Norne Field model with well data and 4D seismic data. For well data (production rates and RFT data), the data in the Norne Field are almost always ‘covered’ by the data when they are examined separately. When pairs of data are examined jointly, however, the well data show large discrepancies. Seismic data showed much greater levels of model deficiency at the lowest level.