6G conditioned spatiotemporal graph neural networks for real time traffic flow prediction

Abstract Accurate, low-latency traffic forecasting is a cornerstone capability for next-generation Intelligent Transportation Systems (ITS). This paper investigates how emerging 6G-era network context specifically per node slice-bandwidth and channel-quality indicators can be fused with spatio-temporal graph models to improve short-term freeway speed prediction while respecting strict real-time constraints. Building on the METR-LA benchmark, we construct a reproducible pipeline that (i) cleans and temporally imputes loop-detector speeds, (ii) constructs a sparse Gaussian-kernel sensor graph, and (iii) synthesizes realistic per-sensor 6G signals aligned with the traffic time series. We implement and compare four model families: Spatio-Temporal GCN (ST-GCN), Graph Attention ST-GAT, Diffusion Convolutional Recurrent Neural Network (DCRNN), and a novel 6G-conditioned DCRNN (DCRNN6G) that adaptively weights diffusion by slice-bandwidth. Our evaluation systematically explores four feature regimes (speeds only; channel quality only; slice bandwidth only; both features), and includes hyperparameter sweeps, ablation studies, and latency profiling on commodity CPUs to reflect edge deployment realities. Empirical results reveal three central findings. First, diffusion-recurrent modeling (DCRNN) produces the best accuracy latency trade-off for large-scale freeway forecasting: it attains test RMSE $$\approx 0.036$$ with average inference latency $$\approx 24$$ ms, comfortably meeting real-time requirements. Second, naïve incorporation of simulated 6G metrics provides only marginal RMSE gains for ST-GCN/ST-GAT and does not improve DCRNN when conditioned simply on bandwidth or CQI; in many cases, small accuracy gains are offset by notable latency penalties. Third, error diagnostics (sensor-wise RMSE, MAE heatmaps, error histograms) expose a small subset of spatially localized hard sensors and episodic time windows that dominate tail errors, indicating where targeted modules (anomaly detectors, incident-aware submodels) could yield outsized improvements. The main contributions of this work are: (1) the first end-to-end benchmarking of 6G-conditioned spatio-temporal GNNs on METR-LA with real-time latency analysis; (2) the introduction and empirical evaluation of a bandwidth conditional diffusion cell (DCRNN6G); and (3) extensive ablation, hyperparameter, and diagnostic studies that quantify both the potential and limitations of network aware fusion for ITS. We conclude by outlining concrete research directions, heterogeneous cross-graph fusion, dynamic adjacency learning, probabilistic forecasting, and real 5G/6G testbed validation that will be critical to realize truly co-optimized transportation and communication systems.