Network Traffic Prediction Based on Fusion of Spatiotemporal Features

Abstract

Accurate traffic prediction is beneficial for enhancing network service quality. However, it is challenging due to multiscale temporal periodic variations and dynamic spatial relationships. Thus, we established a Spatiotemporal Feature Fusion-based Network Traffic Prediction Model (STPM). The temporal prediction module of the model divided the raw data into sequences at the hour, day, and week levels according to the regularities of periodic changes in traffic of different time windows. A combination of two-dimensional convolutions and Long Short-Term Memory (LSTM) networks were used as the encoder and decoder of an attention mechanism to identify the periodic correlations across multiple time scales. The spatial prediction module with a multi-head attention mechanism created grid nodes with interactive relationships under non-Euclidean distances. The module assigned adaptive weights to effectively capture the dynamic correlations between distant grid nodes. These two modules were integrated for traffic prediction. Experiment results showed an $R^{2}$ of 0.8624, demonstrating the effectiveness of STPM.