Abstract: A phased dynamic fault restoration approach based on the entropy-weight TOPSIS for distribution network is given, and it is applied to the design of MAS in this paper. After failure, in order to solve uncertainty problem, DG output and load at each time are forecasted according to the data mining of similar days. In the case of satisfying the constraints, the optimal solution is obtained based on TOPSIS. After applying the optimized strategy, reasonable load shedding in power restoration region can be carried out to make good use of restorable power sources or to cut excessive load in time. In this paper, a three-decker dynamic fault restoration multi-agent system which contains bus agent(BA), state agent(SA) and state subset agent(SSA) is designed. In the end, the feasibility of the design is proved through parsing the typical connection mode of distribution network.