Uncertainty evaluation of reservoir simulation models using particle swarms and hierarchical clustering
Abstract
History matching production data in finite difference reservoir simulation
models has been and always will be a challenge for the industry. The
principal hurdles that need to be overcome are finding a match in the first
place and more importantly a set of matches that can capture the uncertainty
range of the simulation model and to do this in as short a time as possible
since the bottleneck in this process is the length of time taken to run the
model. This study looks at the implementation of Particle Swarm
Optimisation (PSO) in history matching finite difference simulation models.
Particle Swarms are a class of evolutionary algorithms that have shown
much promise over the last decade. This method draws parallels from the
social interaction of swarms of bees, flocks of birds and shoals of fish.
Essentially a swarm of agents are allowed to search the solution hyperspace
keeping in memory each individual’s historical best position and iteratively
improving the optimisation by the emergent interaction of the swarm. An
intrinsic feature of PSO is its local search capability. A sequential niching
variation of the PSO has been developed viz. Flexi-PSO that enhances the
exploration and exploitation of the hyperspace and is capable of finding
multiple minima. This new variation has been applied to history matching
synthetic reservoir simulation models to find multiple distinct history
3
matches to try to capture the uncertainty range. Hierarchical clustering is
then used to post-process the history match runs to reduce the size of the
ensemble carried forward for prediction.
The success of the uncertainty modelling exercise is then assessed by
checking whether the production profile forecasts generated by the ensemble
covers the truth case.