Model Prediction under Uncertainty Using Hierarchical Models

Modelling a physical system such as oil reservoir, however accurate, is subject to uncertainty due to an unrealistic assumption about the model, uncertainty in measured data, and computer model incapability.

A realistic assessment of all sources of uncertainty is a challenging task, especially in oil and gas industry. On the other hand, unrealistic assumptions about model/data can lead to biased estimation of model parameters in a history matching progress. It may also be that the practitioners fail to reliably predict the true model behaviour and oilfield properties in case the uncertainty is not modelled appropriately.

In this paper, we model the uncertainty using two hierarchical models, maximum likelihood model and a full Bayesian hierarchical model. Moreover, we examine the predictive capability of our real reservoir model based on the modelled uncertainty with regards to the true model.

Doing multiple history match trials, a full hierarchical model approach yields better results for our case study than the maximum likelihood approach.