THD‐Net: Temporal Heterogeneity Decoupling for Traffic Flow Forecasting

ABSTRACT

Accurate traffic‐flow forecasting underpins route guidance, signal control, and proactive congestion management, but fine‐resolution traffic series are not governed by a single temporal pattern. Abrupt local disturbances, slower congestion evolution, and recurrent mobility routines contribute differently as the forecast horizon increases, making a homogeneous temporal representation inadequate for direct multihorizon prediction. This study proposes THD‐Net, a temporal‐heterogeneity decoupling network that assigns distinct modeling roles before fusion. A DHA local‐response branch captures recent and weekly lag‐informed dynamics, whereas a subseries Transformer branch abstracts long‐memory interactions and supports trend‐ and recurrence‐oriented summaries; a frequency‐inspired horizon‐profile regularizer constrains the shape of the four‐horizon output. On urban and freeway PeMS scenarios, THD‐Net shows a horizon‐dependent advantage rather than uniform short‐horizon dominance. Strong baselines remain competitive at 5 and 15 min, whereas THD‐Net achieves the lowest five‐seed mean MAE among the compared deep baselines at both 30 and 60 min on both datasets. Paired day‐by‐seed Wilcoxon tests further support the stability of these medium‐ and long‐horizon gains. These results indicate that explicit temporal‐role specialization is a useful inductive bias for robust traffic‐flow forecasting when local fluctuations and slower temporal regularities must be predicted jointly.