Development and evaluation of bidirectional LSTM freeway traffic forecasting models using simulation data.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Long Short-Term Memory (LSTM) models excel at time series prediction.
• Bidirectional LSTMs (BiLSTM) enhance LSTM by processing data in forward and backward directions.
• Accurate short-term traffic forecasting is crucial for managing urban congestion.

Purpose of the Study:

• To develop and evaluate BiLSTM models for short-term traffic forecasting.
• To compare BiLSTM performance against other LSTM models.
• To assess forecasting accuracy under various traffic demand scenarios.

Main Methods:

• Utilized a calibrated micro-simulation model of a Melbourne freeway.
• Generated loop detector data (speed, flow, occupancy) from the simulation.
• Developed and tested multiple LSTM and BiLSTM models for predictions up to 60 minutes ahead.

Main Results:

• BiLSTM models demonstrated superior performance over other predictive models for base year conditions.
• BiLSTM maintained its superior performance across multiple prediction horizons and traffic variables in future year scenarios.
• Forecasting accuracy was evaluated for traffic demand increases of 25-100%.

Conclusions:

• BiLSTM models are highly effective for short-term traffic forecasting on congested freeways.
• The findings support the adoption of BiLSTM for real-world traffic management systems.
• Micro-simulation combined with deep learning offers a robust approach for traffic prediction.