The prediction of time series traffic is a hot issue in machine learning in recent years.It has been found that the prediction accuracy can be greatly improved by approaches of changing the network structure (such as the number of neural network layers, the number of neurons in network layers, the connection mode between network layers, as well as the application of special network layers), and selecting appropriate optimizer and loss function.Here, we propose a multi-layer LSTM (Long Short-Term Memory) algorithm, which is a single model improved on traditional LSTM algorithm, to reduce the model's complexity and improve the efficiency of machine learning.The model includes an input layer, five hidden layers, an output layer, a full connection layer, and also a dropout layer to prevent the machine learning from over-fitting.The model uses adam as optimizer, mlse as loss function, and relu as activation function.The experimental results show that the proposed model has better generalization ability compared with traditional LSTM model.