Citywide Mobile Traffic Forecasting Using Spatial-Temporal Downsampling Transformer Neural Networks

Abstract

The efficient automated network management methods for mobile operators (for example, mobile traffic prediction) are important goals for future mobile networks. Nevertheless, accurately predicting mobile traffic across an entire city is a challenge that requires the consideration of both prediction accuracy and computational complexity, especially in terms of the thousands of regions involved in high-density and complex base station deployment scenarios for 5G/beyond 5G/6G networks. To solve this problem, this study proposed a novel deep learning network based on the transformer, i.e., a spatial-temporal downsampling neural network (STD-Net), which can dynamically and simultaneously exploit the temporal, local, and global spatial dependencies of mobile traffic. To reduce the computational complexity in spatial domains and achieve a balance between generalization (for all regions) and fitness (for each region), the model decomposes a city into patches and focuses on simultaneously exploiting the temporal and local spatial dependencies in each patch via spatial-temporal transformers. This study’s downsampling transformer is responsible for exploiting global spatial dependencies by uniformly sampling mobile traffic throughout all the regions of an entire city. Computing spatial correlations among sampled regions reduces the computational complexity even further. The superior prediction accuracy of the proposed STD-Net model over state-of-the-art baselines was confirmed by experiments on real-world mobile traffic datasets. The effectiveness of each module was tested to further validate the rationale and feasibility of the proposed STD-Net. Analyses of the computational complexity of the STD-Net revealed that its computational cost contains the same quadratic complexity as vanilla transformers.