| Popis: |
Iterative geostatistical history matching is based on the perturbation of reservoir petrophysical properties (i.e. the model parameter space) using stochastic sequential simulation and co-simulation. The convergence of such procedures is ensured by a global optimization, based on the misfit between observed and simulated production data. In these frameworks, the best petrophysical model from the previous iteration is used as secondary variable on the co-simulation of new sets of models, during the next iteration. Normally, this best model corresponds to the model producing the lowest production mismatch from the ensemble generated during the previous iteration. While this procedure ensures convergence along iterations, it also reduces exploration on the model parameter space, which can allow solutions to be trapped in local minima, far from the real solution, thus underestimating existing uncertainties. This work shows how an intelligent selection of models can be used as conditioning data during subsequent iterations in order to overcome these limitations. The selection is based on sampling the model parameter space using linear prediction weighting, derived from the similarity of production curves obtained from multiple realizations on a given iteration. This procedure was implemented in a semi-synthetic reservoir (IPE-I) and the results were compared against a conventional approach. |
