Incorporating Multivariate Auxiliary Information for Traffic Prediction on Highways

Area of Science:

• Intelligent Transportation Systems (ITSs)
• Traffic Engineering
• Machine Learning for Transportation

Background:

• Traffic flow prediction is crucial for effective traffic management in Intelligent Transportation Systems (ITSs).
• Existing models often overlook multivariate auxiliary information and lack interpretability, especially in highway scenarios.
• Predicting traffic flow is complex due to factors like weather and time.

Purpose of the Study:

• To develop a novel traffic prediction model, Multi-variate and Multi-horizon prediction based on Long Short-Term Memory (MMLSTM), for improved highway traffic management.
• To effectively incorporate multivariate auxiliary information (weather, time) and multi-horizon time spans into traffic flow prediction.
• To enhance the interpretability of traffic prediction models.

Main Methods:

• Utilized a multi-horizon bidirectional Long Short-Term Memory (LSTM) model to fuse multivariate auxiliary information across different time spans.
• Integrated an attention mechanism and a multi-layer perceptron for sophisticated traffic flow prediction.
• Employed weather and time data for model interpretability.

Main Results:

• The proposed MMLSTM model demonstrated superior performance compared to baseline methods on traffic prediction tasks.
• Successfully incorporated auxiliary information like weather and time to improve prediction accuracy.
• Achieved better results on the Hangst and Metr-la datasets.

Conclusions:

• MMLSTM offers a significant advancement in highway traffic flow prediction by leveraging multivariate data and multi-horizon analysis.
• The model provides interpretable predictions, aiding traffic management strategies.
• The approach effectively addresses limitations of existing traffic prediction methods.