Bus Single-Trip Time Prediction Based on Ensemble Learning.

Area of Science:

• Transportation Science
• Data Science
• Machine Learning

Background:

• Accurate bus single-trip time prediction is vital for passenger decision-making and efficient bus scheduling.
• Bus operations are influenced by numerous factors, presenting a significant challenge for precise time prediction.
• Bus trip times exhibit distinct nonlinear and seasonal patterns.

Purpose of the Study:

• To enhance the accuracy of bus single-trip time prediction.
• To propose and evaluate a data-driven framework for bus trip time prediction.
• To compare the performance of various base and ensemble prediction models.

Main Methods:

• Utilized five base prediction algorithms: Long Short-term Memory (LSTM), Linear Regression (LR), K-Nearest Neighbor (KNN), Extreme Gradient Boosting (XGBoost), and Gate Recurrent Unit (GRU).
• Developed three ensemble models using Random Forest (bagging), AdaBoost (boosting), and Linear Regression (stacking).
• Proposed a three-phase data-driven framework: traffic data analysis, feature extraction, and ensemble model prediction.

Main Results:

• Ensemble models demonstrated superior prediction accuracy compared to individual base models.
• The Random Forest ensemble model achieved the highest prediction accuracy among the tested ensemble methods.
• Linear Regression (LR) provided better prediction accuracy than the other four base models (LSTM, KNN, XGBoost, GRU).

Conclusions:

• Ensemble methods significantly improve bus trip time prediction accuracy.
• Random Forest is an effective ensemble technique for this prediction task.
• While ensemble models outperform base models, Linear Regression shows strong performance among individual algorithms.