In order to resolve the issues of prematurity and instability of catastrophic genetic algorithm （CGA）, an improved CGA （ICGA） is proposed, and an improved catastrophic operator related to the generation number is de- signed. Moreover, considering both the global performance and the convergence speed, two probability algorithms respectively for the crossover related to the generation number and for the mutation related to the fitness are de- signed. The proposed ICGA is finally applied to the reactive power optimization of the IEEE 14-bus and the IEEE 30-bus systems. The results show that ICGA is applicable to the reactive power optimization of power system due to its good global performance and high convergence speed.