A hybrid car-following behavior prediction framework for snowy weather based on stacking ensemble learning
- 1Changsha Technology Investment Management Co., Ltd., Changsha, China.
- 2National Engineering Research Center of Highway Maintenance Technology, Changsha University of Science and Technology, Changsha, China.
- 3School of Traffic and Transportation Engineering, Central South University, Changsha, China.
Abstract
To improve the accuracy and robustness of vehicle car-following behavior prediction on snowy highways, this paper proposes a hybrid prediction framework based on Stacking ensemble learning. The framework integrates a theoretically driven improved Intelligent Driver Model (IDM) with two state-of-the-art data-driven models—the Prediction-enhanced Intelligent Driver Model (P-IDM) and a Transformer-LSTM hybrid network—within a Stacking ensemble architecture. First, an Icy Condition Enhanced IDM (ICES-IDM) model is developed by introducing a road adhesion coefficient mapping function and a speed perception error coefficient to parameterize the classical IDM for snowy environments. Second, a P-IDM model is constructed by incorporating a self-attention-based preceding vehicle speed prediction module into the classical IDM framework to enhance responsiveness to sudden maneuvers of the leading vehicle. Third, a Transformer-LSTM dual-stream network (TL-Net) is designed to jointly predict car-following acceleration by parallel extraction of temporal attention features and recurrent neural network features. Within the Stacking framework, ICES-IDM and P-IDM serve as physics-driven base learners, TL-Net serves as a data-driven base learner, and XGBoost is adopted as the meta-learner to construct a hybrid acceleration prediction model. The model is trained and validated using naturalistic driving data collected in Norway—comprising a dataset of 311,908 observations that link vehicle measurements with weather, road surface, and lighting conditions, along with a five-month road record dataset from Narvik. Experimental results demonstrate that the proposed model achieves a root mean square error (RMSE) in acceleration prediction under snowy conditions that is approximately 28.1% lower than that of the standalone ICES-IDM model, 9.2% lower than that of the TL-Net model, and 4.2% lower than that of the conventional Stacking-IDM-LSTM framework. The model exhibits excellent generalization capability across varying snow intensity levels, providing theoretical support for vehicle car-following safety warning systems in snowy environments.
Related
- StackHPpred: A Stacking-based ensemble learning framework for the identification of peptide hormones using multi-view feature representations
- An ensemble framework for robust drug response prediction and large-scale sensitivity profiling across cancers
- SeverYX: An ensemble framework for robust variant severity classification
- Spatial Prediction of Opisthorchis viverrini Infection Risk Using Multi-Source Remote Sensing, Hexagonal Grid Optimization, and Ensemble Machine Learning: A Case Study From Northeastern Thailand.
- SIMORGH: Ensemble-Aware Geometric Deep Learning for Apo-State and Cryptic Ligand Binding Site Prediction
BibTeX
@article{luo2026hybrid,
title = {{A hybrid car-following behavior prediction framework for snowy weather based on stacking ensemble learning}},
author = {Zengjie Luo and Fan Geng and Zhaolei Zhang},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361759},
url = {https://doi.org/10.1371/journal.pone.0361759},
note = {CC BY},
}
RIS
TY - JOUR
TI - A hybrid car-following behavior prediction framework for snowy weather based on stacking ensemble learning
AU - Zengjie Luo
AU - Fan Geng
AU - Zhaolei Zhang
PY - 2026
DA - 2026/10/30
JO - PLOS One
DO - 10.1371/journal.pone.0361759
UR - https://doi.org/10.1371/journal.pone.0361759
AB - To improve the accuracy and robustness of vehicle car-following behavior prediction on snowy highways, this paper proposes a hybrid prediction framework based on Stacking ensemble learning. The framework integrates a theoretically driven improved Intelligent Driver Model (IDM) with two state-of-the-art data-driven models—the Prediction-enhanced Intelligent Driver Model (P-IDM) and a Transformer-LSTM hybrid network—within a Stacking ensemble architecture. First, an Icy Condition Enhanced IDM (ICES-IDM) model is developed by introducing a road adhesion coefficient mapping function and a speed perception error coefficient to parameterize the classical IDM for snowy environments. Second, a P-IDM model is constructed by incorporating a self-attention-based preceding vehicle speed prediction module into the classical IDM framework to enhance responsiveness to sudden maneuvers of the leading vehicle. Third, a Transformer-LSTM dual-stream network (TL-Net) is designed to jointly predict car-following acceleration by parallel extraction of temporal attention features and recurrent neural network features. Within the Stacking framework, ICES-IDM and P-IDM serve as physics-driven base learners, TL-Net serves as a data-driven base learner, and XGBoost is adopted as the meta-learner to construct a hybrid acceleration prediction model. The model is trained and validated using naturalistic driving data collected in Norway—comprising a dataset of 311,908 observations that link vehicle measurements with weather, road surface, and lighting conditions, along with a five-month road record dataset from Narvik. Experimental results demonstrate that the proposed model achieves a root mean square error (RMSE) in acceleration prediction under snowy conditions that is approximately 28.1% lower than that of the standalone ICES-IDM model, 9.2% lower than that of the TL-Net model, and 4.2% lower than that of the conventional Stacking-IDM-LSTM framework. The model exhibits excellent generalization capability across varying snow intensity levels, providing theoretical support for vehicle car-following safety warning systems in snowy environments.
SN - 1932-6203
ER -