In order to extend the application by using supervised learning methods in non-intrusive load identification, a load identification method based on Recurrent Neural Networks (RNN) model is proposed. Firstly, the time window for detecting load event is introduced, and then the harmonic components are taken as the load characteristics to be used as the inputs of the RNN model. According to the memory of its memory history input feature quantity, the internal mapping of the input to the output as well as the RNN load identification method for time series inputs are established. Furthermore, in this model, the suitable activation function and loss function are selected in order to avoid the "gradient disappearance" problem. Finally, the experiment on the single and multi-load identification demonstrates the model can effectively realize the identification of the load status. This work is supported by Science and Technology Project of China Southern Power Grid Co., Ltd. (No. 090000KK52160049).