The methodology presented in this paper is a two-stage optimization approach that can be applied to large system level models, in this case using a Stochastic Petri Net (SPN) framework, to produce an equivalent model response at a reduced computational cost. The method consists of generating a reduced SPN which approximates the behavior of its large counterpart with a shorter simulation time. Parameters in this reduced structure are updated following a combined Approximate Bayesian Computation and Subset Simulation framework. In the first stage, optimization of the reduced model via a Genetic Algorithm provides a first approximation of the optimal solutions for the full system level model. In the second stage, these approximate optimal solutions then form the starting point of a short optimization of the large SPN to fine tune the results using a reduced solution space. This method is demonstrated for a sub-section of an SPN of a fire protection system. Optimization of the full model with a Genetic Algorithm is compared to the optimization through this two-stage approach to demonstrate the capability of the methodology. Results show good model agreement at a reduced computational cost.