Empirical study of the effects of physics-guided machine learning on freeway traffic flow modelling: model comparisons using field data

Recent studies have shown the successful implementation of classical model-based approaches (e.g. macroscopic traffic flow modelling) and data-driven approaches (e.g. machine learning – ML) to model freeway traffic patterns, while both have their limitations. Even though model-based approaches could depict real-world traffic dynamics, they could potentially lead to inaccurate estimations due to traffic fluctuations and uncertainties. In data-driven models, the acquisition of sufficient high-quality data is required to ensure the model performance. However, many transportation applications often suffer from data shortage and noises. To overcome those limitations, this study aims to introduce and evaluate a new model, named as physics-guided machine learning (PGML), that integrates the classical traffic flow model (TFM) with the machine learning technique. This PGML model leverages the output of a traffic flow model along with observational features to generate estimations using a neural network framework. More specifically, it applies physics-guided loss functions in the learning objective of neural networks to ensure that the model not only consists with the training set but also shows lower errors on the known physics of the unlabelled set. To illustrate the effectiveness of the PGML, this study implements empirical studies with a real-world dataset collected from a stretch of I-15 freeway in Utah. Experimental study results show that the proposed PGML model could outperform the other compatible methods, including calibrated traffic flow models, pure machine learning methods, and physics unguided machine learning (PUML).