Abstract
Accurate traffic prediction is fundamental to effective urban mobility. This paper investigates the shift from mono-class to multi-class traffic modelling and its implications for physics-guided machine learning techniques. We examine two approaches for multi-class highway traffic prediction, encompassing both passenger vehicles and trucks: (i) a physics-regularized Gaussian Process (GP) model that integrates constraints from the multi-class METANET model, and (ii) a Long Short-Term Memory (LSTM) network that incorporates these constraints into its loss function. Beyond evaluating prediction accuracy, computational cost, and implementation complexity, we highlight the challenges introduced by multi-class formulations—particularly the need to account for diverse vehicle dynamics. Our findings provide insights into the trade-offs and feasibility of each approach and assess the impact of transitioning from mono-class to multi-class models on predictive performance.

This publication has 42 references indexed in Scilit: