Traffic Matrix Prediction Based on Multilevel Discrete Wavelet Transform Network and LSTM

Abstract

Accurate prediction of future traffic matrix (TM) for communication networks can help network managers adjust traffic scheduling policies in advance, which can reduce the probability of link congestion and improve the efficiency of network operation. This paper proposes a TM prediction framework based on multilevel discrete wavelet transform network and long and short-term memory neural network (MDWTN-LSTM). Discrete wavelet transform (DWT) is introduced into TM prediction to extract multi-scale time-frequency features, which can help the neural network model to grasp traffic trends. And then we approximately realized the DWT scheme through the linear layer in the neural network, so that the wavelet transform is embedded in the neural network in a tightly coupled form and participates in the training of model parameters, finally achieves the effect of global parameter optimization and improves both prediction accuracy and adaptability of the prediction framework. The MDWTN-LSTM based model is verified by a variety of benchmarks using the real-world data sets, the experimental results show that the proposed framework can achieve relatively superior prediction accuracy. And compared with the theoretical optimal result, 98.6% and 91.1% of the maximum link utilization bias for traffic scheduling based on MDWTN-LSTM is less than 10%, which is sufficient to support reliable traffic engineering.