Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting

Lingxiao Cao, Bin Wang, Guiyuan Jiang, Yanwei Yu, Junyu Dong

TL;DR

STDN tackles traffic forecasting by jointly modeling dynamic high-order spatio-temporal interactions and trend–seasonality structure. It introduces a dynamic relationship graph, spatio-temporal embeddings, and a spatio-temporally aware decomposition, followed by a GRU encoder and Transformer-based decoder with Bottleneck Transformer blocks; the approach yields $O(TD + LND)$ encoder/decoder complexity and $O((T+N)D + N^3)$ embedding complexity, with preprocessing of Laplacian eigen decomposition. Empirical results on PeMS04, PeMS07, and JiNan show state-of-the-art accuracy and improved efficiency, and ablations confirm the essential roles of dynamic graphs, spatio-temporal embeddings, and the STD module. The JiNan dataset release expands evaluation coverage to inner-city dynamics, enhancing the practical relevance of traffic forecasting research for ITS applications.

Abstract

Traffic prediction is critical for optimizing travel scheduling and enhancing public safety, yet the complex spatial and temporal dynamics within traffic data present significant challenges for accurate forecasting. In this paper, we introduce a novel model, the Spatiotemporal-aware Trend-Seasonality Decomposition Network (STDN). This model begins by constructing a dynamic graph structure to represent traffic flow and incorporates novel spatio-temporal embeddings to jointly capture global traffic dynamics. The representations learned are further refined by a specially designed trend-seasonality decomposition module, which disentangles the trend-cyclical component and seasonal component for each traffic node at different times within the graph. These components are subsequently processed through an encoder-decoder network to generate the final predictions. Extensive experiments conducted on real-world traffic datasets demonstrate that STDN achieves superior performance with remarkable computation cost. Furthermore, we have released a new traffic dataset named JiNan, which features unique inner-city dynamics, thereby enriching the scenario comprehensiveness in traffic prediction evaluation.

Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting

TL;DR

STDN tackles traffic forecasting by jointly modeling dynamic high-order spatio-temporal interactions and trend–seasonality structure. It introduces a dynamic relationship graph, spatio-temporal embeddings, and a spatio-temporally aware decomposition, followed by a GRU encoder and Transformer-based decoder with Bottleneck Transformer blocks; the approach yields encoder/decoder complexity and embedding complexity, with preprocessing of Laplacian eigen decomposition. Empirical results on PeMS04, PeMS07, and JiNan show state-of-the-art accuracy and improved efficiency, and ablations confirm the essential roles of dynamic graphs, spatio-temporal embeddings, and the STD module. The JiNan dataset release expands evaluation coverage to inner-city dynamics, enhancing the practical relevance of traffic forecasting research for ITS applications.

Abstract

Traffic prediction is critical for optimizing travel scheduling and enhancing public safety, yet the complex spatial and temporal dynamics within traffic data present significant challenges for accurate forecasting. In this paper, we introduce a novel model, the Spatiotemporal-aware Trend-Seasonality Decomposition Network (STDN). This model begins by constructing a dynamic graph structure to represent traffic flow and incorporates novel spatio-temporal embeddings to jointly capture global traffic dynamics. The representations learned are further refined by a specially designed trend-seasonality decomposition module, which disentangles the trend-cyclical component and seasonal component for each traffic node at different times within the graph. These components are subsequently processed through an encoder-decoder network to generate the final predictions. Extensive experiments conducted on real-world traffic datasets demonstrate that STDN achieves superior performance with remarkable computation cost. Furthermore, we have released a new traffic dataset named JiNan, which features unique inner-city dynamics, thereby enriching the scenario comprehensiveness in traffic prediction evaluation.

Paper Structure

This paper contains 17 sections, 14 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Problem Definition
  4. Methodology
  5. Module 1: Dynamic Relationship Graph Learning
  6. Module 2: Spatio-Temporal Embeddings Learning
  7. Module 3: Trend-Seasonality Decomposition
  8. Encoder-Decoder Architecture
  9. Experiments
  10. Datasets
  11. Baseline Methods
  12. Evaluation Metrics and Experimental Settings
  13. Performance Comparison
  14. Ablation Study
  15. Parameter Sensitivity Study
  16. ...and 2 more sections

Figures (4)

  • Figure 1: The overview of the proposed framework. MLP: multi-layer perceptron, GCN: graph convolution network.
  • Figure 2: Parameter sensitivity study on PeMS04 and PeMS07 datasets.
  • Figure 3: The computational time cost on PeMS04 and JiNan datasets.
  • Figure 4: The MAE on the validation part of PeMS04 dataset during the training process.

Theorems & Definitions (2)

  • Definition 1: Traffic Network
  • Definition 2: Traffic Forecasting