A Review of Traffic Flow Prediction Research from the Perspective of Physics-Informed Fusion — Based on Physics-Informed Graph Neural Networks

Abstract

Among the principal functions of intelligent transportation systems, traffic flow prediction stands out as a critical one. Purely data-driven methods face the bottleneck of insufficient generalization capability in scenarios involving sparse data, noise interference, and anomalous events, while classical traffic flow theory is constrained by difficulties in parameter calibration and the oversimplification of model assumptions. The rise of physics-informed neural networks (PINNs) offers a new theoretical framework for the deep integration of these two types of methods. This paper systematically reviews the evolution of traffic prediction methods from traditional time series models to graph neural networks, thoroughly analyzes the theoretical foundations and numerical implementations of physics-based traffic flow models, and surveys the current research status of physics-informed neural networks in the transportation domain. On this basis, the paper identifies the limitations of existing research in terms of task coverage, model architecture, and analytical dimensions, and envisions future research directions from three perspectives: dynamic graph learning, embedding higher-order physical models, and the comparison of hard versus soft constraints. This study demonstrates that physical constraints, as inductive biases, can effectively narrow the hypothesis space of models, providing a methodological reference for interdisciplinary integration in traffic flow prediction.