CA-DTNet: a trace-driven communication-aware digital twin network for robust high-resolution traffic forecasting under heterogeneous V2X conditions

Reliable high-resolution traffic forecasting is essential for connected intelligent transportation systems, where observations are delivered through heterogeneous vehicular networks. Most forecasting models assume clean, immediately available measurements, although operational inputs can be stale or incomplete. This paper proposes CA-DTNet, a communication-aware digital twin network for robust short-term traffic forecasting. One-second pNEUMA traffic states are combined with ordered CICV5G and SEE-V2X trace segments through a deterministic replay protocol. Packet loss directly produces sample-and-hold observations and increasing Age of Information (AoI), while latency, jitter, load, throughput, and radio-quality measurements condition a learned reliability gate. CA-DTNet uses a shared spatiotemporal representation for clean-state reconstruction and quantile forecasts of speed, vehicle count, and flow at 5-, 10-, and 30-second horizons. Across three seeds, the model obtains the lowest mean absolute error and weighted absolute percentage error in 24 of 27 domain–horizon–target combinations. Its reconstruction $$R^{2}$$ exceeds 0.985 in both communication domains. Exact matched-window analysis yields a mean absolute cross-domain MAE shift of 0.20% and a significant domain effect in only one of nine forecasting tasks. With 71,203 trainable parameters and approximately 0.393 ms inference latency per sample, CA-DTNet improves forecasting accuracy without relying on greater capacity than DCRNN or Graph WaveNet. The results support trace-driven, uncertainty-aware traffic digital twins while also delimiting the current evidence to one traffic source and two replay domains.