Abstract: Power quality problems caused by a large quantity of nonlinear loads connected to the distribution network are increasingly serious,and the accuracy of the nonlinear load modeling,to a certain extent,affects the harmonic power flow calculation and power quality analysis of the distribution network.Given that it is difficult to describe the nonlinear load with the mechanism dynamic model under complex operation conditions,and it is difficult to avoid the modeling error based on the prediction method,a dual-layer recurrent neural network model is adopted in this paper to model nonlinear loads,which includes the preliminary power prediction layer and error correction layer of the recurrent neural network. According to the training samples of load power and voltage,the preliminary power prediction layer predicts the load power of the next moment. In the error correction layer of the recurrent neural network,according to the deviation between the predicted power of the previous layer and measured power,the predicted power of the next moment is corrected by feedbacks. As the bus voltage fluctuation affects the load power,the STL algorithm is adopted to decompose the time-series bus voltage,and the residual component threshold is set to determine whether the error correction layer of the neural network is activated. The model result shows that proposed modeling method can better realize the fitting of nonlinear loads,and avoid the decline of modeling accuracy when bus voltage fluctuates greatly.