Data integration and uncertainty evaluation for large scale automatic history matching

The goal of this research is to develop a theoretical basis and practical technology for the integration of production data and geological data in a way which provides an accurate assessment of the uncertainty in reservoir description and reservoir performance prediction.; The adjoint method and LBFGS optimization algorithm make large scale history matching possible. However, computational experiments reveal that the original implementation of the adjoint method and the LBFGS optimization algorithm may encounter the following problems: (i) sometimes, the adjoint method generates incorrect sensitivities and gradient information; (ii) the optimization algorithm often does not give as good a match of production data as is expected; (iii) on rare occasions, the LBFGS converges to a model which results in over/undershooting problems. In this dissertation, the author analyzes the main reasons that cause these problems and investigates practical ways to overcome them.; First, we modify the adjoint equations used for calculating the gradient of the objective function or sensitivity of data for a case where the initial condition is dependent on some model parameters.; Second, several different strategies are investigated to improve the efficiency of the optimization algorithm: (1) a more robust line search algorithm motivated by the theoretical result that the Wolfe conditions should be satisfied; (2) balancing the data integrated by applying damping procedures; (3) application of constraints on the permeability/porosity fields; and (4) adding auxiliary terms that force all data to be sensitive to model parameters.; Finally, we apply the randomized maximum likelihood (RML) method with the improved optimization algorithm to quantify the uncertainty of rock properties and future performance prediction by history matching multi-phase flow production data for the well-known PUNQS3 example. The results show that the efficiency of history matching is much improved. In most cases, the objective function can converge to the expected level, the estimated model is consistent with the prior model and the performance prediction is acceptable. The uncertainty in the future performance prediction assessed by the RML method is similar to that by the EnKF method. Both uncertainties in model parameters and future performance predictions are decreased significantly by history matching the production data.