An Empirical Comparison of Urban Road Travel Time Prediction Methods—Deep Learning, Ensemble Strategies and Performance Evaluation

Urban road travel time prediction is a critical component of intelligent transportation systems. However, accurate prediction remains challenging due to the high volatility and randomness of urban road traffic. This study compares the predictive performance of deep learning methods with traditional shallow learning methods for urban road travel time prediction, and explores the potential for improving prediction effectiveness through multi-run training strategies for hyperparameter optimization and ensemble learning. The study employs urban road travel time data from Shenzhen, China, designing 10 prediction scenarios with different sliding time window lengths and prediction scales. To address the randomness issue in deep learning models, we adopt a strategy of conducting five independent training runs for each hyperparameter configuration and using statistical measures to obtain stable performance evaluations. For model comparison, we implement multiple shallow learning models (linear regression, ridge regression, LASSO regression, random forest) and deep learning models (LSTM-DNN, DNN), and construct ensemble learning models based on linear regression and LSTM-DNN as base learners. The experimental results demonstrate that: (1) deep learning models generally outperform shallow learning models in terms of Mean Absolute Percentage Error (MAPE), particularly the LSTM-DNN model which achieves the best MAPE values across all prediction scenarios with 30 min sliding time windows; (2) in terms of Root Mean Square Error (RMSE), shallow learning models such as random forest perform better in most scenarios; (3) ensemble learning models show certain advantages in some prediction scenarios, but the improvement effects are limited and scenario-dependent. The research results provide empirical reference for method selection in urban road travel time prediction, while also revealing the limitations of single data sources and the differences in model performance under different evaluation metrics, offering directions for future research.