Blending measurements and numerical models: a novel methodological approach for land subsidence prediction with uncertainty quantification

Abstract

The use of numerical models for land subsidence prediction above producing hydrocarbon reservoirs has become a common and well-established practice since the early '90s. Usually, uncertainties in the deep rock behavior, which can affect the forecast capability of the models, have been taken into account by running multiple simulations with different constitutive laws and mechanical properties. Then, the most uncertain parameters were calibrated to reproduce available subsidence measurements. The objective of this work is to propose a novel methodological approach for land subsidence prediction and uncertainty quantification by integrating the available monitoring information in numerical models using ad hoc Data Assimilation techniques. The proposed approach allows to: (i) train the model with the available data and improve its accuracy as new information comes in, (ii) quantify the prediction uncertainty by providing confidence intervals and probability measures instead of deterministic outcomes, and (iii) identify the most appropriate rock constitutive model and geomechanical parameters. The methodology is tested in synthetic models of production from hydrocarbon reservoirs. The numerical experiments show that the proposed approach is a promising way to improve the effectiveness and reliability of land subsidence models.