MSTL-GLTP: A Global–Local Decomposition and Prediction Framework for Wireless Traffic

Abstract

With the rapid development of the Internet of Things and increasingly rigid communication requirements, the wireless traffic prediction framework is experiencing a transition from edge/cloud server deployment to edge–cloud collaborative deployment. However, it remains a significant challenge to balance prediction accuracy and overall complexity based on edge–cloud collaboration networks. In this article, we propose a multiple seasonal-trend decomposition using loess-based global–local traffic prediction (MSTL-GLTP) framework that assures prediction accuracy while maintaining low complexity. Specifically, we first decompose the cellular traffic into the multiseasonal, trend, and residual components through the MSTL algorithm. Subsequently, multiseasonal components are clustered and fed into the bidirectional long short-term memory (Bi-LSTM) model to capture global tendency. Meanwhile, we exploit a distance-assisted attention mechanism to minimize global loss. Besides, a local network module consisting of the temporal convolutional network (TCN) and Gaussian process regression (GPR) model is deployed in the edge devices to learn the dynamic regional and local traffic. The experimental results demonstrate that MSTL-GLTP outperforms the state-of-the-art baselines by capturing global–local spatiotemporal correlation and achieves accuracy and complexity equilibrium when predicting wireless traffic.