Non-intrusive load identification can achieve the flexible bilateral interaction between power grid and users, which is of great significant in the development of smart grid. Neural network was frequently employed in non-intrusive load identification because of its self-learning ability and low computation complexity. In order to overcome the shortcomings that BP neural network traps into local optima easily and has a low convergence speed, this paper proposes a new method based on General Regression Neural Network (GRNN). Firstly, this method uses transient features such as power, harmonic and switch time as the inputs of GRNN. Secondly, neural network structure is constructed based on Parzen non-parametric estimation method. Thirdly, simulated annealing algorithm is adopted to get the best smoothing parameter. Finally, RGNN network model is built to identify the load. Experimental results have proved that the proposed method has higher identification accuracy and training speed than BP neural network.