Edge-Side Cellular Network Traffic Prediction Based on Trend Graph Characterization Network

Predicting edge-side cellular network traffic stands as a pivotal facilitator for network automation in next-generation communication systems. However, the traffic data at the edge exhibits significant heterogeneity, inhomogeneity, and volatility due to geographic location, human activities, and demand diversification, thus making accurate network traffic prediction a rigorous challenge. To solve this problem, this paper proposes a novel cellular network traffic prediction model in the edge-managed multi-base station (BS) scenarios, named trend graph characterization network (TGCN). Structurally, TGCN has three key components of trend feature extractor, temporal feature extractor and predictor. Firstly, the high-dimensional trend feature of traffic can be captured by the combination of ordinal pattern transition network (OPTN) and graph attention network (GAT). Furthermore, in the temporal feature extractor neural circuit policy (NCP) is introduced for multi-scale time-varying dependent features. Finally, a fully-connected layer serves as the approximator of BS traffic. On real-world datasets, we verify the superiority of our proposal via statistical analysis, prediction accuracy and ablation experiments.