Data-driven approaches for modeling train control models: Comparison and case studies.

Area of Science:

• Railway Engineering
• Control Systems
• Machine Learning

Background:

• Automatic train control (ATC) algorithms degrade due to environmental factors and equipment wear.
• Traditional train control models rely on mechanical analysis, requiring costly field trials for validation.

Purpose of the Study:

• To develop data-driven approaches for accurate train control modeling.
• To overcome the limitations of traditional mechanical train control models.
• To improve the efficiency and accuracy of train control algorithms in railway systems.

Main Methods:

• Collected three years of explicit train operation data from Beijing Metro.
• Developed three data-driven models: linear regression (LAM), nonlinear regression (NRM), and deep neural network (DNN).
• Customized DNN architecture, including input/output layers and batch normalization, for railway-specific characteristics.
• Used LAM and NRM as benchmarks to evaluate DNN performance.

Main Results:

• The deep neural network (DNN) model demonstrated significantly enhanced prediction accuracy compared to LAM and NRM.
• Customized DNN network structure improved training efficiency and predictive capabilities.
• Data-driven approaches were successfully validated using field data.

Conclusions:

• Data-driven models, particularly the customized DNN, offer a superior alternative to traditional mechanical models for train control.
• The developed approaches provide a cost-effective and accurate method for designing efficient train control algorithms.
• Successful application in Beijing Metro demonstrates the practical viability of these data-driven techniques in real-world railway systems.