Service composition provides a flexible way to quickly enable new application functionalities in next generation networks. Composite services are turned to be more and more complex, and the performance analysis of composite service is an important issue for service providers who want to get better QoS. In this paper, we proposed a Stochastic Petri Net based approach for modeling web service composition constructs and calculating the equivalent delay time of the composite web service. Four basic constructs: sequence, concurrent, choice and loop constructs are mapped into Petri Net. Methods to calculate equivalent delay time are illustrated in detail. A complex composite service can be reduced by repeatedly replacing basic constructs within it with equivalent single transition. This approach can be applied in performance analysis for complex composite web service, as well as complex systems and workflow.