Machine Learning Model Application and Comparison in Actuated Traffic Signal Forecasting.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning Applications
• Traffic Engineering

Background:

• Traditional traffic signal prediction methods are inadequate for dynamic, traffic-actuated systems.
• Direct communication in intelligent traffic systems poses significant security risks.
• The increasing availability of data makes machine learning a viable solution for traffic signal forecasting.

Purpose of the Study:

• To investigate the efficacy of machine learning models for traffic signal forecasting.
• To address the limitations of traditional methods in dynamic traffic environments.
• To develop a secure and accurate traffic signal prediction system.

Main Methods:

• Collected real-world traffic data from an intersection using IoT-enabled detectors.
• Trained and compared several machine learning models: Baseline, Dense, Linear, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM).
• Evaluated LSTM performance on a larger dataset and as a binary classifier.

Main Results:

• LSTM achieved a test accuracy exceeding 95% for traffic signal forecasting.
• The median deviation in LSTM predictions was as low as 2 seconds.
• LSTM demonstrated high performance as a binary classifier with over 92% accuracy and an AUC close to 1.

Conclusions:

• Long Short-Term Memory (LSTM) networks are highly effective for accurate traffic signal forecasting.
• Machine learning, particularly LSTM, offers a secure and reliable alternative to traditional communication-based prediction methods.
• The developed LSTM model significantly improves the performance of intelligent traffic systems.