Today's production planning faces major challenges: The life cycle of new products gets noticeably shortened while development and production costs rise. This demands increasingly complex production systems are required. To minimize costs, these systems must be well understood and controlled. Simulation offers an extensive degree of understanding and controlling of production systems. However, such a utilization of simulation requires detailed models. While current models are already very detailed, the desire for even more detailed simulation models continues. Even with the ever increasing computing power, this hunger for more detailed and widespread simulation models cannot be satisfied. To solve this problem, the complexity of the simulation model can be reduced. This reduction process is normally done by the modeling engineer who must have a certain degree of experience to create valid reduced models. Because the work of such an engineer is very expensive in terms of time and monetary costs, automated approaches have been proposed. Most of the automated approaches rely on observed data from several cost-intensive simulation runs. The gathered data is only valid for observed conditions. Using this data in unknown conditions can be erroneous. Furthermore, most automated approaches are restricted to very specific simulation models. The concept of this thesis is an automated approach that overcomes several of the mentioned drawbacks. Instead of processing every token individually, several tokens are handled in the same manner as one reference token. The presented method doesn't need a cost-intensive preprocessing step to gather data. Instead, the analysis of the system's behavior is constantly redone to adapt the behavior of the coarsening to new conditions. Due to this highly dynamic approach, the presented concept is capable of adapting to changing conditions at runtime.