科研之友
微信
新浪微博
Facebook
分享链接
摘要
History matching has been applied over the years to validate reservoir models using production history data. Assisted history matching techniques utilized are numerous and their effectiveness is dependent on the nature of the history matching problem. For problems involving the definition of global reservoir parameters (e.g. rock compressibility) as inputs, optimizers are usually selected to explore the search space for one or more optimal solutions that fit production history. This approach is sufficient for cases where essential reservoir behaviour is captured in the model. However, in situations where the model constructed omits important reservoir physics, matching using optimizers that perturb only global parameters becomes futile.
On the other hand optimization methods, especially those capable of being applied at grid cell level, give reliable history matching results by changing static grid properties (e.g. porosity, permeability, net-to-gross ratio) over their uncertainty limits. In an unconventional approach, grid-block-based optimization methods can also be used to screen models for omitted reservoir behaviour after a lengthy and unyielding history matching process.
In this article, the capabilities of optimization techniques like direct search (Covariance Matrix Adaptation Evolution Strategy), proxy modeling (Artificial Neural Network) and gradient-based (Adjoint) algorithm coupled with an experimental design method to explore models for omitted reservoir behaviour at grid level is assessed extensively. A simple synthetic model with a unique geologic scenario is constructed with model solution utilized as observation data. Permeability in all three directions is defined as an uncertainty parameter for all optimizers at grid block level. Performance criteria such as the accuracy of static property map replicated, computational cost and required time are presented. Results obtained from the application of all three optimization techniques are compared.
- 出版日期2016-12
