A comparative study and simple baseline for travel time prediction.

Area of Science:

• Transportation Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate travel time prediction (TTP) is crucial for traffic management and trip planning.
• Existing TTP methods are complex, involving multiple stages, making it hard to identify key factors influencing accuracy.

Purpose of the Study:

• To investigate the impact of various TTP process steps on prediction accuracy.
• To compare different data imputation and feature engineering techniques.
• To evaluate base models against hybrid models for TTP.

Main Methods:

• Evaluated data imputation techniques (deep learning, interpolation, max value).
• Assessed the influence of temporal features and weather conditions.
• Compared five hybrid TTP models and base models (LSTM, XGBoost) using real-world datasets from Taiwan and California.

Main Results:

• Data preprocessing, including feature engineering, significantly impacts TTP accuracy.
• Base models (LSTM, XGBoost) demonstrated superior performance over hybrid models on real-world data.
• A novel fusion model combining XGBoost and LSTM via a gating network achieved enhanced prediction accuracy.

Conclusions:

• Data preprocessing is a critical determinant of TTP accuracy.
• Simpler base models can be more effective than complex hybrid models.
• The proposed XGBoost-LSTM fusion model offers a robust and accurate approach to TTP.