Related Experiment Videos
Sensor-Driven Short-Term Forecasting on the Metropolitan LA Traffic Dataset: A Comparative Study for Multi-Step
Bowen Dong1, Xinyu Zhang2, Weiyan Zhu3
1School of Electrical Automation and Information Engineering, Tianjin University, Tianjin 300072, China.
Sensors (Basel, Switzerland)
|June 26, 2026
Summary
Understanding sensor data issues is key for accurate short-term traffic forecasting. This study introduces a diagnostic framework and a new hybrid model, GETFormer, to improve intelligent transportation systems.
Area of Science:
- Intelligent Transportation Systems
- Machine Learning
- Data Science
Background:
- Short-term traffic forecasting is vital for intelligent transportation systems.
- Deep learning models for traffic forecasting face challenges due to sensor data characteristics like zero-value prevalence and heterogeneity.
- Existing research lacks systematic analysis of how these data properties impact model performance and guide model selection.
Purpose of the Study:
- To systematically analyze sensor data characteristics and their impact on deep learning model performance for traffic forecasting.
- To develop a diagnostic framework for understanding architecture-specific failure modes in traffic sensing.
- To propose a novel hybrid architecture, GETFormer, for improved traffic forecasting.
Main Methods:
- Applied a sensor-network diagnostic framework to the METR-LA dataset (207 inductive loop detectors, 5-min resolution).
- Benchmarked four representative architectures: Transformer, Spatio-Temporal Graph Convolutional Network (STGCN), Diffusion Convolutional Recurrent Neural Network (DCRNN), and Gated Temporal Convolutional Network (Gated TCN).
- Conducted a per-sensor regression analysis linking zero-value ratios to model-specific prediction errors.
Main Results:
- Identified specific sensor data characteristics (zero-value prevalence, heterogeneity, correlations) that cause architecture-specific failures.
- Quantitatively linked higher zero-value ratios to increased prediction errors in certain models.
- Developed Graph-Enhanced Transformer (GETFormer), a hybrid model outperforming others in specific conditions.
Conclusions:
- Sensor data properties significantly influence the performance of deep learning models in traffic forecasting.
- A diagnostic approach is crucial for evidence-based model selection in real-world intelligent transportation systems.
- The proposed GETFormer architecture offers a promising direction for developing robust urban traffic sensing models.
