Based on the concept of measurement uncertainty, the robust state estimation method for power system with maximum normal measurement rate (MNMR) has good identification ability of bad data. However, the model is difficult to solve. In existing researches, the model is approximated and solved by the modern interior point method, but the problems such as lower identification effect due to approximation are existed. Therefore, based on the state estimation model of MNMR, a hybrid particle swarm optimization (PSO) algorithm with hybrid mutation is used to propose a bad data identification algorithm based on graphics processing unit (GPU) parallel acceleration. Without approximating the MNMR model and according to the architecture characteristics of the GPU parallel compution, the algorithm designs a parallel acceleration strategy combining coarse and fine granularity. The results of the case analysis show that the proposed algorithm has low false detection rate and missed detection rate for bad data, and has good identification ability of bad data, short calculation time and high acceleration efficiency, which could meet the actual operation requirements.