Hybrid model predicts car-following on snowy Norwegian highways more accurately

Predicting how a driver will speed up or slow down behind another vehicle matters for road safety, and snow makes it harder. In a study published in the peer-reviewed journal PLoS ONE, researchers from Central South University and Changsha University of Science and Technology describe a hybrid framework that blends physics-based models of driving with neural networks. According to the study, the combination predicts car-following acceleration on snowy roads more accurately than its parts do individually, and somewhat better than an earlier hybrid design. It is an example of traditional traffic theory and machine learning being combined rather than one replacing the other.

Car-following models describe how a vehicle responds to the one in front of it. A classical example is the Intelligent Driver Model, or IDM, a set of equations that sets a driver's acceleration from the traffic situation ahead. The authors first adapt it to winter. Their version, ICES-IDM, adds a mapping function that ties the model to the road adhesion coefficient, a measure of how much grip tyres have on the surface, and a coefficient representing errors in how drivers perceive speed. A second model, P-IDM, keeps the IDM structure but adds a module that forecasts the speed of the leading vehicle using self-attention, a technique that lets a network weigh which past observations matter most. The aim, the authors say, is to respond better when the car ahead manoeuvres suddenly.

The third component is purely data-driven. TL-Net runs two kinds of neural network in parallel: a Transformer, the attention-based architecture that underpins modern language models, and an LSTM (long short-term memory network), an older type of recurrent network that processes a sequence step by step while carrying a running memory. Features from both streams are used jointly to predict acceleration.

The three models are then combined through stacking, an ensemble technique in which several base models each make a prediction and a second model, the meta-learner, learns how best to merge them. Here ICES-IDM and P-IDM act as the physics-driven base learners, TL-Net as the data-driven one, and the meta-learner is XGBoost, a widely used method built on gradient-boosted decision trees. The framework was trained and validated on naturalistic driving data from Norway, meaning recordings of ordinary driving rather than staged experiments: a dataset of 311,908 observations linking vehicle measurements with weather, road surface and lighting conditions, plus a five-month road record dataset from Narvik.

The main measure is root mean square error, or RMSE, a standard gauge of average prediction error that penalises large misses more heavily than small ones. The authors report that under snowy conditions the hybrid model's RMSE for acceleration is about 28.1% lower than that of ICES-IDM on its own, 9.2% lower than TL-Net alone and 4.2% lower than a conventional stacking framework combining an IDM with an LSTM. They also report that it generalises well across different levels of snow intensity.

The authors present the result as theoretical support for car-following safety warning systems in snow, not as a working system. The gain over the earlier stacking approach is modest, and because the data come from Norway, the abstract leaves open how the framework would fare on other roads, in other climates or in an operational warning system.