Reservoir characterization and forecasting require construction of reliable and history matched reservoir models. In gradient-based automatic history matching procedure, most of the CPU time is spent on the sensitivity calculation. When the adjoint algorithm is used to calculate the sensitivity, the CPU time is approximately proportional to the magnitude of observed data. If we can identify observed data with most information content, we can minimize the history matching time using adjoint method and hence make the history matching faster.
A methodology has been developed in this work to identify the important production data and discard those "less useful" data in the automatic history matching procedure. This enables us to save significant computational time without sacrificing much of the history matching result. Our approach ensures that the selected observed data contains little redundancy, and gives similar inverted result based on the contents of all the observed data set.
The method is validated by conducting history matches for synthetic data sets. The comparisons of computational time and the quality of the inverted results indicate that this method is promising and practical for the automatic history matching. Using this approach, computational effort can be reduced by as much as ten times without sacrificing the quality of the results.
The final form of the automatic history matching procedure is to minimize an objective function which includes the information for both a priori distribution and likelihood function. Typically, the minimization of the objective function utilizes gradient information and therefore requires a procedure for generating sensitivity coefficients.1,2,3 This is the most time consuming part and forms the core part of the automatic history matching. In this work, a new attempt was made to improve the efficiency of automatic history matching by reducing both the number of observed data and the number of model parameters used in the inversion procedure to estimate reservoir parameters.
In a realistic case, the observed data are usually provided with certain errors and possibly a large amount of data points. It is not feasible to use all the production data in the automatic history matching because it is time consuming. It is intuitive that some observed production data contain more information about reservoir properties than other data. Then the question comes up: how to select those higher quality observed data out from the whole set of production data so that the history matching procedure will not be compromised without the use of all the observed data?
With these motivations, the relationship between the production data and the reservoir rock properties is studied. A data selection method has been developed in this work based on the SVD (Singular Value Decomposition) method. The SVD method provides us a way to select the optimal observed data based on the decomposed components of the sensitivity matrix.
To make the automatic history matching efficient, the data selecting method is combined with the powerful scale-splitting scheme. This enables us to significantly save the computational time by reducing the number of the observed data as well as the model parameters. Several synthetic cases using the data selection methods are shown in this study. The results show that the optimal observed data can be successfully selected by using SVD method, and the number of observed data can be significantly reduced without losing the quality of inverted permeability in reservoir model.