Multi-scale network traffic prediction based on attention mechanism and long short-term memory network

This paper proposes a hybrid model that combines empirical mode decomposition(EMD), attention mechanism(AM), and long short-term memory network(LSTM) for accurate traffic prediction. The EMD technique is applied to decompose the network traffic data into intrinsic mode functions(IMFs) and a residual error with similar characteristics. These components are reconstructed and combined to form feature vectors, which serve as input for the LSTM network. An AM is integrated to capture essential temporal information and focus on significant temporal features. The final traffic prediction is obtained by aggregating the individual predictions from each component. Experimental results demonstrate that the proposed hybrid model surpasses traditional autoregressive integrated moving average, support vector machine(SVM), recurrent neural network(RNN), independent LSTM, and LSTM-AM models in terms of prediction accuracy.