Real-Time Traffic Turning Flow 2026
This research investigates forecasting traffic turning movements when little or no local training data is available. It evaluates TimesFM, a decoder-only time-series model, and combines it with Empirical Mode Decomposition and Random Forest to incorporate temporal patterns and external factors.
- Published in Artificial Intelligence for Transportation, Volume 5, 2026, article 100046.
- Forecasting in few-shot and zero-shot settings with limited historical data.
- Comparison with Temporal Fusion Transformer, DeepAR, and LSTM.
- Evaluation across weather conditions and forecast horizons, including temporal alignment of predictions.
76 Waterloo intersections for training and testing with limited local data; 36 Milton intersections for testing without local training data.
Two forecasting approaches: a separate encoder and decoder (top), and a decoder-only model that predicts each next step directly (bottom).
Traffic forecasting workflow, from data preparation to model evaluation.
Three TimesFM forecasting variants: baseline, decomposition, and decomposition with Random Forest covariates.
- Intelligent Transportation
- Time-Series Forecasting
- TimesFM
- Few-Shot Learning
- Zero-Shot Learning
Ce Zhang, Yuzhe You, Guangyuan Pan, Matthew I. Muresan, Liping Fu