Memory guided adaptive spatiotemporal forecasting for city level cellular traffic

Accurate prediction of city-level cellular network traffic is essential for proactive resource allocation, congestion avoidance, and quality-of-service management in next-generation mobile networks, including under the ongoing evolution toward IPv6. Existing spatiotemporal forecasting methods face two limitations: graph-based approaches rely on pre-defined or statically learned adjacency structures that may not adequately reflect time-varying inter-cell dependencies, and most architectures lack explicit mechanisms to encode recurring traffic patterns at both global network-wide and local per-cell granularities. To address these limitations, we propose CellMem, a memory-guided adaptive spatiotemporal network that combines a DLinear backbone, dual-branch prototype memory, a prototype-driven dynamic graph, and a prototype-aware gated residual. Its core novelty is an end-to-end shared- \(W_g\) coupling: the same global prototype attention weights retrieve pattern-specific correction features and mix learnable low-rank sub-graphs into a sample-specific adjacency matrix. On the Milan Telecom Italia dataset under one-step-ahead forecasting, CellMem reduces Internet RMSE from \(155.17 ± 1.68\) (SA-GCN) to \(143.05 ± 0.57\) and improves \(R^2\) from 0.9579 to 0.9642, with consistent gains on SMS and Call traffic. Source code is available at https://doi.org/10.5281/zenodo.21835202 .