Application of Machine Learning Techniques in Short-term Travel Time Prediction Using Multiple Data Sources

Taghipour, Homa; Parsa, Amir Bahador; Mohammadian, Abolfazl · 2020 · Crossref

DOI: 10.31224/osf.io/k3nq6

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the limitations of current highway travel time prediction methods, which often rely on naïve techniques and limited data sources, resulting in unreliable estimates that can delay travelers. The authors aim to improve short-term travel time prediction by integrating multiple data sources—including loop detectors, probe vehicles, weather conditions, accidents, road works, and special events—to account for diverse factors influencing traffic. The research focuses on a 5-mile segment of the westbound Eisenhower Highway in Chicago, utilizing data from April to December 2017. The methodology employs two machine learning algorithms: K-Nearest Neighbors (KNN) and Random Forest. The study first processes raw data from inductive loop detectors, probe vehicles (via the NPMRDS dataset), and other sources through rigorous cleaning and imputation techniques, including temporal, spatial, and historical estimation to handle missing or erroneous records. These datasets are combined into a link-based dataset aggregated in 5-minute intervals. The models are trained to predict harmonic average speed, which serves as a proxy for travel time, across prediction horizons ranging from immediate estimation to 60 minutes ahead in 5-minute increments. The dataset is split into 80% for training and 20% for testing. Results indicate that both models perform well in short-term prediction, but the Random Forest model consistently outperforms KNN. At the moment of prediction, Random Forest achieved an R-squared score of 95.3%, compared to 92.9% for KNN. While accuracy decreased as the prediction horizon extended, Random Forest remained more robust, achieving 84% accuracy at 60 minutes ahead versus 76.8% for KNN. Feature importance analysis revealed that traffic variables, particularly occupancy and vehicle count, were the most significant predictors. Temperature was the next most influential factor, while accidents, road works, and special events had lower direct impact, likely due to their infrequency and correlation with traffic variables. Sensitivity analysis confirmed a linear relationship between increased occupancy and reduced speed, whereas count and annual average daily traffic showed nonlinear effects on speed. The study concludes that Random Forest is the superior method for short-term travel time prediction, particularly for horizons of 15 minutes or less. It highlights the critical role of occupancy data from loop detectors in predicting travel time and demonstrates that incorporating weather data improves model performance. The findings suggest that integrating diverse data sources with advanced machine learning techniques can significantly enhance the reliability of travel time information for both road users and traffic engineers.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction