Abstract
Automatic history matching may be used to condition reservoir simulation models by including time-lapse seismic data. Stochastic optimization algorithms are used to ensure a good search of the parameter space to ensure proper determination of the best models. These approaches can require many thousands of simulations for large dimensional problems. Divide and conquer is an assisted history matching approach that enables deconvolution of the parameters so that they can be searched more efficiently and also leads to better uncertainty analysis.
We present an application of this approach on the Nelson field. Nine years of production history data are used along with baseline and monitor surveys. Localised variations were made to permeability and net:gross. The reservoir was divided into separate parameter regions by combining experimental design and proxy model analysis. The former enabled insignificant parameters to be discarded while the latter showed that each region could be treated as a separate history matching sub-problem. Each sub-problem was then solved simultaneously using an adapted stochastic genetic algorithm.
The results showed that a forty-two dimensional problem could be reduced to a combination of three 9D problems and a 3D problem due to the spatial deconvolution of parameters and misfits. An improved match was obtained for the production and seismic data. Compared to a full stochastic search of the parameter space, the number of models was several orders of magnitude smaller. Further, improved uncertainly analysis was made possible resulting in better understanding of the future behaviour of the reservoir.
An improved match to reservoir models leads to better confidence in their prediction and thus they can be used more effectively in reservoir management. The method presented here to improve the match retains the benefits of stochastic searching without the penalty of requiring an impractical number of simulations.
We present an application of this approach on the Nelson field. Nine years of production history data are used along with baseline and monitor surveys. Localised variations were made to permeability and net:gross. The reservoir was divided into separate parameter regions by combining experimental design and proxy model analysis. The former enabled insignificant parameters to be discarded while the latter showed that each region could be treated as a separate history matching sub-problem. Each sub-problem was then solved simultaneously using an adapted stochastic genetic algorithm.
The results showed that a forty-two dimensional problem could be reduced to a combination of three 9D problems and a 3D problem due to the spatial deconvolution of parameters and misfits. An improved match was obtained for the production and seismic data. Compared to a full stochastic search of the parameter space, the number of models was several orders of magnitude smaller. Further, improved uncertainly analysis was made possible resulting in better understanding of the future behaviour of the reservoir.
An improved match to reservoir models leads to better confidence in their prediction and thus they can be used more effectively in reservoir management. The method presented here to improve the match retains the benefits of stochastic searching without the penalty of requiring an impractical number of simulations.
| Original language | English |
|---|---|
| Pages | 1-10 |
| Number of pages | 10 |
| DOIs | |
| Publication status | Published - Jun 2012 |
| Event | SPE Europec/EAGE Annual Conference - Copenhagen, Denmark Duration: 4 Jun 2012 → 7 Jun 2012 |
Conference
| Conference | SPE Europec/EAGE Annual Conference |
|---|---|
| Country/Territory | Denmark |
| City | Copenhagen |
| Period | 4/06/12 → 7/06/12 |