Focusing on the power decreasing of large-scale applications in network-on-chip, this paper proposed amodified genetic algorithm based method on low-power mapping. With communication weights of task nodes and structural features of mapping platform, this method acquired better initial mapping solution set with the consideration of task node priority and its connection. Moreover, we introduced the roulette wheel selection, best-neighbor selection and reverse evolution, and selected the initial solution with a certain probability at each iteration to prevent the algorithm stagnation.Experimental results show that, when maintaining the same task model and mapping platform, compared with the genetic algorithm and random mapping algorithm, our proposed algorithm greatly decreases the energy consumption.