Multiscale Traffic Dynamics Representation for Forecasting via MEMD-Guided Dual-Branch Recurrent Networks.

Area of Science:

• Transportation Science
• Data Science
• Signal Processing

Background:

• Traffic flow forecasting is complex due to mixed temporal patterns like trends and fluctuations.
• Existing methods struggle to capture both long-term trends and short-term dynamics effectively.

Purpose of the Study:

• To propose a novel Multivariate Empirical Mode Decomposition (MEMD)-guided dual-branch recurrent framework for multistep traffic flow forecasting.
• To enhance the accuracy of traffic flow predictions by effectively modeling different temporal patterns.

Main Methods:

• Utilized MEMD for alignment-preserving multivariate decomposition into frequency-aligned components.
• Reconstructed components into low-frequency trend and high-frequency residual.
• Employed Long Short-Term Memory (LSTM) for trend modeling and Bidirectional Gated Recurrent Unit (Bi-GRU) for residual modeling.
• Integrated predictions using a lightweight fusion head.

Main Results:

• Achieved competitive performance on PeMS04 and PeMS08 datasets.
• Reported MAE, RMSE, and MAPE of 19.67/31.59/12.95% on PeMS04 and 15.51/24.43/9.86% on PeMS08.
• Demonstrated relative performance gains of up to 5.89% on PeMS04 and 5.35% on PeMS08 compared to baselines.

Conclusions:

• MEMD-guided trend-residual representation learning significantly improves multistep traffic flow forecasting.
• The proposed dual-branch framework effectively captures both long-term and short-term traffic dynamics.
• The method offers a promising approach for accurate and reliable traffic flow prediction.