A multi-agent system (MAS) based approach for modeling and simulation of power system restoration is presented in this paper. The power system restoration process is represented as a multi-agent system with hierarchical architecture. Two major kinds of agents are included, which are management agent (MGAG) and practical components in power system (PCAG). PCAG includes three types of agents: generator agent (GAG), substation agent (SAG) and load agent (LAG). When a subsystem with necessary components such as GAGs, SAGs and LAGs is built up, one MGAG is generated to in charge of the subsystem and provides negotiations between different PCAGs and with other subsystems. Each agent is granted with its own knowledge base, database and decision support base. In order to figure out a feasible restoration plan, a distributed computing structure is proposed. The computation tasks such as constraints check of operations and path search for power transmission are assigned to every agent in the MAS. Both decentralized control and centralized control are employed in this model to make up for the disadvantages of each control method. A sample simulation in a MAS platform is provided to demonstrate the validity of this model