Spatiotemporal Prediction of Wireless Network Traffic Based on Multi-Task Deep Learning

Accurate wireless network traffic prediction has become an urgent need and major challenge in network management, driven by rapid technological advances and data-driven strategies. Currently, state-of-the-art algorithms typically neglect spatio-temporal feature extraction, resulting in limited accuracy. Therefore, this paper proposes a spatiotemporal and temporal sequence joint prediction framework based on multi-granularity feature sharing (MTSN). Rooted in multi-task learning, MTSN models traffic's spatio-temporal and multi-granularity temporal features to achieve effectively prediction. Specifically, we first construct a multi-granularity data partitioning layer that forms multi-scale data association matrices through temporal dimension decomposition, enabling collaborative optimization of cross-granularity traffic prediction tasks. Then, 3D Convolutional Networks (Conv3D) Conv3D extracts spatio-temporal correlations while Kolmogorov-Arnold Networks with Gated Recurrent Units (KAN-GRU) captures multi-granularity temporal dependencies. These features are fused to create a joint representation en-compassing both global trends and local details. Finally, fully connected networks extract granularity-specific unique features, enhancing overall prediction. Using a real Italian telecommunications dataset, MTSN demonstrates effectiveness through comparative analyses of prediction performance, trend fitting, statistical distribution, and spatial distribution consistency. Experimental results demonstrate that our model achieves significant performance improvements in key metrics such as RMSE and MAE compared to advanced models including STGCN and T-GCN.