The ensemble Kalman filter (EnKF) has gained increased popularity for history matching and continuous reservoir-model updating. It is a sequential Monte Carlo approach that works with an ensemble of reservoir models. Specifically, the method uses cross covariance between measurements and model parameters estimated from the ensemble. For practical field applications, the ensemble size needs to be kept small for computational efficiency. However, this leads to poor approximations of the cross covariance and can cause loss of geologic realism from unrealistic model updates outside the region of the data influence and/or loss of variance leading to ensemble collapse. A common approach to remedy the situation is to limit the influence of the data through covariance localization. In this paper, we show that for three-phase-flow conditions, the region of covariance localization strongly depends on the underlying flow dynamics as well as on the particular data type that is being assimilated, for example, water cut or gas/oil ratio (GOR). This makes the traditional distance-based localizations suboptimal and, often, ineffective. Instead, we propose the use of waterand gas-phase streamlines as a means for covariance localization for water-cutand GOR-data assimilation. The phase streamlines can be computed on the basis of individual-phase velocities which are readily available after flow simulation. Unlike the total velocity streamlines, phase streamlines can be discontinuous. We show that the discontinuities in water-phase and gas-phase streamlines naturally define the region of influence for water-cut and GOR data and provide a flow-relevant covariance localization during EnKF updating. We first demonstrate the validity of the proposed localization approach using a waterflood example in a quarter-five-spot pattern. Specifically, we compare the phase streamline trajectories with cross-covariance maps computed using an ensemble size of 2,000 for both water-cut and GOR data. The results show a close correspondence between the time evolution of phase streamlines and the cross-covariance maps of water-cut and GOR data. A benchmark uncertainty quantification (the PUNQ-S3) (Carter 2007) model application shows that our proposed localization outperforms the distance-based localization method. The updated models show improved forecasts while preserving geologic realism.
[1]
J. W. Barker,et al.
Quantifying Uncertainty in Production Forecasts: Another Look at the PUNQ-S3 Problem
,
2001
.
[2]
G. Evensen.
Data Assimilation: The Ensemble Kalman Filter
,
2006
.
[3]
Behnam Jafarpour,et al.
Reservoir Characterization With the Discrete Cosine Transform
,
2009
.
[4]
Dean S. Oliver,et al.
THE ENSEMBLE KALMAN FILTER IN RESERVOIR ENGINEERING-A REVIEW
,
2009
.
[5]
Geir Nævdal,et al.
An Iterative Ensemble Kalman Filter
,
2011,
IEEE Transactions on Automatic Control.