One of the main concern in the O&G business is generating reliable production profile forecasts. Such profiles are the cornerstones of optimal technico-economical management decisions. A workflow combining different methodologies to integrate and reduce most of the subsurface uncertainties using multiple history matched models (explaining the past) to infer reasonably reliable production forecasts is proposed.
Using experimental design theory, a sensitivity study is first performed to scan the whole range of static and dynamic uncertain parameters using a proxy-model of the fluid flow simulator. Only the most sensitive ones with respect to an objective function (quantifying the mismatch between the simulation results and the observations) are retained for subsequent steps.
Assisted History Matching tools are then used to get multiple History matched models, an order of magnitude faster than traditional History Matching processes. Updated uncertain parameters (selected from the sensitivity studies) may be picked anywhere in the direct problem building workflow.
Using the Bayesian framework, a posterior distribution of the most sensitive parameters are derived from the a priori distributions and a non-linear proxy model of the likelihood function. The later is computed using experimental design, kriging and dynamic training techniques.
Multiple History Matched models together with a posteriori parameter distributions are finally used in a joint modeling approach to capture the main uncertainties and to obtain typical (P10-P90) probabilistic production profiles.
This workflow has been applied to a gas storage real case submitted to significant seasonal pressure variations. Probabilistic operational pressure profiles for a given period can then be compared to the actual gas storage dynamic behaviour to assess the added value of the proposed workflow.
Introduction
Getting probabilistic production forecasts of a reservoir through a risk analysis is closely linked to uncertainty quantification 1. Uncertainty quantification should end with a posteriori uncertain parameter distributions, reflecting all the knowledge that we have on the reservoir and explaining as much as possible the observation data. The uncertainties may result both from the observation data quality as well as from the numerical modeling steps. On top of that, the fluid flows equations we are dealing with are non linear, which leads to possible complex hydrodynamical behaviour. CPU constraints make inefficient brute force Monte Carlo sampling of the a posteriori distributions from a priori ones. Most of the current approaches try to scan more efficiently (in a given timeframe) the uncertain parameters space using reliable proxy models 2 of the fluid flow simulator and/or to take advantage of new computational power 3,4,5,6.
Due to the numerous uncertain parameters involved in the history matching process, a screening of the parameters space is first performed (i.e. a sensitivity study) to retain only the most sensitive ones with respect to the history matching criteria, possibly reducing the variation intervals. This is a pre-processing step of the history matching job.
