Optimizing travel time reliability with XAI: A Virginia interstate network case using machine learning and meta-heuristics

Abstract

This paper applies machine learning models to predict travel time reliability in transportation networks, using XGBoost, LightGBM, and CatBoost optimized with seven metaheuristic algorithms. The models were fine-tuned with a four-year dataset (2014–2017) covering 59 interstate sections in Virginia. Key features Link Length, AADT/mile/lane, Total Rate, and PRCP/1000 were identified as influential factors for travel time index prediction. Results revealed that XGBoost optimized with Grey Wolf Optimizer (GWO) achieved the highest accuracy at 92 %, surpassing the base model. LightGBM-GWO and CatBoost-GWO also demonstrated improvements, scoring up to 89 %. GWO outperformed other optimization methods, delivering superior accuracy with fewer control parameters. Feature importance analysis highlighted Link Length and AADT/Lane.mile as critical predictors. This research enhances travel time reliability prediction, providing insights for transportation planning and management. Future work includes exploring multi-objective optimization and integrating additional features to refine model accuracy further.