Kahrobaei, Siavash (Delft University of Technology) | Habibabadi, M. Mansoori (Delft University of Technology) | Joosten, Gerard J. P. (Sharif University of Technology) | Van den Hof, Paul M. J. (Shell Global Solutions International) | Jansen, Jan-Dirk (Eindhoven University of Technology)
Classic identifiability analysis of flow barriers in incompressible single-phase flow reveals that it is not possible to identify the location and permeability of low-permeability barriers from production data (wellbore pressures and rates), and that only averaged reservoir properties in between wells can be identified. We extend the classic analysis by including compressibility effects. We use two approaches: a twin experiment with synthetic production data for use with a time-domain parameter-estimation technique, and a transfer-function formalism in the form of bilaterally coupled four-ports allowing for an analysis in the frequency domain. We investigate the identifiability, from noisy production data, of the location and the magnitude of a low-permeability barrier to slightly compressible flow in a 1D configuration. We use an unregularized adjoint-based optimization scheme for the numerical time-domain estimation, by use of various levels of sensor noise, and confirm the results by use of the semianalytical transfer-function approach. Both the numerical and semianalytical results show that it is possible to identify the location and the magnitude of the permeability in the barrier from noise-free data. By introducing increasingly higher noise levels, the identifiability gradually deteriorates, but the location of the barrier remains identifiable for much-higher noise levels than the permeability. The shape of the objective-function surface, in normalized variables, indeed indicates a much-higher sensitivity of the well data to the location of the barrier than to its magnitude. These theoretical results appear to support the empirical finding that unregularized gradient-based history matching in large reservoir models, which is well-known to be a severely ill-posed problem, occasionally leads to useful results in the form of model-parameter updates with unrealistic magnitudes but indicating the correct location of model deficiencies.