The choice of a 'Best' assisted history matching algorithm
conference paper
Computer-assisted history matching is the act of systematicalty adjusting a ‘prior’ reservoir model using measured data until
its simulated production response closely reproduces the past behavior of the reservoir. Thereafler, the updated, ‘posterior’,
model is expected to predict future reservoir behavior with a higher degree of confidence, particularly if the adjustments are
constrained by known geological properties of the reservoir. The accuracy of the history matching process depends on the
quality of the prior reservoir model and the quality and quantity of measured data, typically production data (pressures and
flow rates at well locations). The uncertainties in the prior reservoir model and in the measurements have a strong effect on the
uncertainty in the parameter values of the posterior model, and thus on the predictions to be made with that model. A common
way to include these uncertainties is through the specification of spatial covariance matrices for prior reservoir parameters and
measurement errors respectivety. Here we consider two assisted history matching techniques that use such a covariance-based
uncertainty description, the representer method and the ensemble Kalman filter, both originating from oceanography. In their
most simple form, used for state estimation in linear systems, they lead to identical estimates. However, when applied to
nonlinear reservoir models their resuits differ. We compare the quatity of the parameter estimates using both methods for a
synthetic ‘twin experiment’ where the true parameter field is known altowing for comparison with the estimates. We consider
simple two-dimensional water flooding cases with ‘true’ reservoir permeabilities represented by either a Gaussian random
field or a channelized geology. The corresponding spatial covariance matrices are obtained with the aid of geostatistical
ensembles of prior models, conditioned on known permeability values in the wells. We compare the quality of the history
match, the permeability estimates, and the predicted water rates after the history matching period using both methods, starting
from two different priors for each of the methods. We conclude, for the cases considered in this study, 1) that because of the
limited information present in production data assessing the ‘best’ method to reconstruct the parameter fietd is not possible; 2)
that from the point of view of prediction after the assimilation period, the ‘best’ assisted history matching method appeared to
be the representer method, but that this might be related to the particular implementation of the methods in our study, and 3)
that further research is required into the effect of treating nonlinearities in the system and output equations, and especially into
the effect of iterative procedures.
its simulated production response closely reproduces the past behavior of the reservoir. Thereafler, the updated, ‘posterior’,
model is expected to predict future reservoir behavior with a higher degree of confidence, particularly if the adjustments are
constrained by known geological properties of the reservoir. The accuracy of the history matching process depends on the
quality of the prior reservoir model and the quality and quantity of measured data, typically production data (pressures and
flow rates at well locations). The uncertainties in the prior reservoir model and in the measurements have a strong effect on the
uncertainty in the parameter values of the posterior model, and thus on the predictions to be made with that model. A common
way to include these uncertainties is through the specification of spatial covariance matrices for prior reservoir parameters and
measurement errors respectivety. Here we consider two assisted history matching techniques that use such a covariance-based
uncertainty description, the representer method and the ensemble Kalman filter, both originating from oceanography. In their
most simple form, used for state estimation in linear systems, they lead to identical estimates. However, when applied to
nonlinear reservoir models their resuits differ. We compare the quatity of the parameter estimates using both methods for a
synthetic ‘twin experiment’ where the true parameter field is known altowing for comparison with the estimates. We consider
simple two-dimensional water flooding cases with ‘true’ reservoir permeabilities represented by either a Gaussian random
field or a channelized geology. The corresponding spatial covariance matrices are obtained with the aid of geostatistical
ensembles of prior models, conditioned on known permeability values in the wells. We compare the quality of the history
match, the permeability estimates, and the predicted water rates after the history matching period using both methods, starting
from two different priors for each of the methods. We conclude, for the cases considered in this study, 1) that because of the
limited information present in production data assessing the ‘best’ method to reconstruct the parameter fietd is not possible; 2)
that from the point of view of prediction after the assimilation period, the ‘best’ assisted history matching method appeared to
be the representer method, but that this might be related to the particular implementation of the methods in our study, and 3)
that further research is required into the effect of treating nonlinearities in the system and output equations, and especially into
the effect of iterative procedures.
TNO Identifier
522359
Source title
Paper preseented at the SPE EUROPEC/EAGE Annual Conference and Exhibition held in Barcelona, Spain, 14-17 June 2010
Pages
1-15
Files
To receive the publication files, please send an e-mail request to TNO Repository.