Traffic agencies require current and near‑future speeds for every link in their network, yet fixed sensors only cover a fraction of it. This work tackles joint nowcasting and one‑hour forecasting of hourly speeds on all 4,127 directed freeway links in the Dallas‑Fort Worth area using sensors on just 643 links (15.6%). Historical probe‑vehicle data supervise every link during training, while inference relies solely on sensor readings – a setting we call probe‑supervised sparse sensing.
We introduce GCRN+, a graph convolutional recurrent network that predicts deviations from a historical speed profile. Its key components are: (1) a parameter‑free, direction‑aware kernel interpolation that produces sensor anomaly features and serves as an output baseline; (2) a diffusion‑convolutional GRU operating on the directed road graph; (3) a low‑rank adaptive graph; and (4) learnable link embeddings. The model scales linearly with network size and can be trained on a CPU. Sensor readings are emulated from probe speeds, isolating spatio‑temporal inference from detector calibration.
On four held‑out test days, GCRN+ achieves an RMSE of 2.21 mph for the current hour and 2.39 mph for the next hour on unsensed links, outperforming the historical average (3.23 mph). The best baselines that use the same interpolation features reach 2.53 mph and 2.75 mph. Recent models without these features (STID, STAEformer, ImputeFormer, MoGERNN) attain at most 2.54 mph and 2.53 mph. GCRN+ is best across all distance bands except links disconnected from any sensor in the graph. Ablation studies attribute most of the gain to the link embeddings and the adaptive graph, which can be trained thanks to probe labels for every link.
For congestion detection on unsensed links, recall improves from 77% (historical average) and 86% (best competitor) to 89% while precision stays at 0.94.
Review