Deep Learning Ensemble Approach for Predicting Expected and Confidence Levels of Signal Phase and Timing Information...

Area of Science:

• Intelligent Transportation Systems
• Machine Learning for Traffic Management
• Deep Learning in Automotive Applications

Background:

• Accurate prediction of Signal Phase and Timing (SPaT) is crucial for optimizing traffic flow and enabling advanced driver-assistance systems.
• Existing deep learning methods have limitations in predicting SPaT information and associated confidence levels.
• The integration of SPaT prediction into systems like Green Light Optimal Speed Advisory (GLOSA) and Eco-Cooperative Adaptive Cruise Control (Eco-CACC) requires enhanced predictive capabilities.

Purpose of the Study:

• To propose and evaluate a novel transformer encoder-based architecture for improved SPaT information and confidence level prediction.
• To compare the performance of the proposed architecture against traditional deep learning methods (MLP, LSTM, CNNLSTM).
• To assess the effectiveness of model consensus as an indicator of prediction confidence.

Main Methods:

• Development of an ensemble of predictors combining a transformer encoder with MLP, LSTM, and CNNLSTM.
• Data-driven prediction of SPaT information from traffic signal controllers at six intersections.
• Evaluation of three tasks: predicting phase change within 20s, predicting exact change time within 20s, and assigning a confidence level.

Main Results:

• The transformer encoder architecture outperformed existing deep learning methods in predicting phase changes and exact change times.
• Achieved 96% average accuracy for predicting phase changes within 20 seconds.
• Demonstrated an average Mean Absolute Error (MAE) of 1.49s for predicting exact change times, outperforming other models.
• Showcased that model consensus effectively indicates prediction confidence, with high-consensus predictions being accurate 90.2% of the time within one second.

Conclusions:

• Transformer encoder-based architectures offer superior performance for SPaT prediction tasks compared to conventional deep learning models.
• Ensemble methods leveraging model consensus provide a reliable measure of prediction confidence, crucial for real-world applications.
• The proposed approach enhances the reliability and performance of intelligent transportation systems like GLOSA and Eco-CACC.