HyperD: Hybrid Periodicity Decoupling Framework for Traffic Forecasting

TL;DR

HyperD (Hybrid Periodic Decoupling), a novel framework that decouples traffic data into periodic and residual components, achieves state-of-the-art prediction accuracy, while offering superior robustness under disturbances and improved computational efficiency compared to existing methods.

Abstract

Accurate traffic forecasting plays a vital role in intelligent transportation systems, enabling applications such as congestion control, route planning, and urban mobility optimization. However, traffic forecasting remains challenging due to two key factors: (1) complex spatial dependencies arising from dynamic interactions between road segments and traffic sensors across the network, and (2) the coexistence of multi-scale periodic patterns (e.g., daily and weekly periodic patterns driven by human routines) with irregular fluctuations caused by unpredictable events (e.g., accidents, weather, or construction). To tackle these challenges, we propose HyperD (Hybrid Periodic Decoupling), a novel framework that decouples traffic data into periodic and residual components. The periodic component is handled by the Hybrid Periodic Representation Module, which extracts fine-grained daily and weekly patterns using learnable periodic embeddings and spatial-temporal attention. The residual component, which captures non-periodic, high-frequency fluctuations, is modeled by the Frequency-Aware Residual Representation Module, leveraging complex-valued MLP in frequency domain. To enforce semantic separation between the two components, we further introduce a Dual-View Alignment Loss, which aligns low-frequency information with the periodic branch and high-frequency information with the residual branch. Extensive experiments on four real-world traffic datasets demonstrate that HyperD achieves state-of-the-art prediction accuracy, while offering superior robustness under disturbances and improved computational efficiency compared to existing methods.

Figures (7)

• Figure 1: Illustration of traffic flow in the PEMS04 dataset. (a) Traffic flow for sensors 54, 73, and 299 over one week. (b) Traffic flow of sensor 54 over three consecutive weeks. Both images demonstrate the strong presence of periodic patterns in traffic data.
• Figure 2: Overview of HyperD, which comprises three main components: (1) The Hybrid Periodic Representation Module encodes daily and weekly embeddings using the Spatial-Temporal Attentive Encoder and generates hybrid periodic patterns. (2) The Frequency-Aware Residual Representation Module encodes the residual using a Spatial-Temporal Frequency Encoder and combines it with the periodic component to yield the final prediction. (3) The Dual-View Alignment Loss separates the prediction into low- and high-frequency parts, which are then aligned with the periodic and residual branches, respectively.
• Figure 3: Robustness testing in the PEMS03 datasets.
• Figure 4: Visualization of the daily and weekly embeddings learned by HyperD, as well as the daily embedding from CycleNet-D and the weekly embedding from CycleNet-W, in the PEMS03 and PEMS04 datasets.
• Figure 5: Detailed architectures of the STAE and the STFE.