| dc.description.abstract | Model Error in probabilistic history matching is an important topic to study, but calculating the model error is a challenge since the truth is uncertain. In this thesis, sources of model error will be discussed briefly; a novel approach has been proposed, first to define model error by using a high-quality model instead of the truth, then statistical parameters of model error will be calculated, and these parameters will be used to account for model error in EnKF.
Two cases have been tested with this approach. In the first case, decline curve analyses were used to model the production rate. Model error has been calculated and accounted for in updating the model with EnKF. The second case studied model error in a 2D reservoir for the upscaling process. The reservoir has been upscaled, and statistical parameters of model error were obtained to be used in updating the model with EnKF.
Results from these two examples showed the importance of the model error in data assimilation. In both cases, it has been proven that neglecting model error caused biases and overconfidence in the forecasted updated model. Additionally, the proposed approach could mitigate the biases and the overconfidence in the forecasted updated model.
The proposed method is in an early stage, and further study should be done to verify and improve it. Other sources of error should be examined. A proper machine learning algorithm could improve the quality of this method to account for model error. | |
