A Traffic Model Integrating Long Short-Term Memory Networks with Multi-class Macroscopic Equations

Accurate traffic state estimation is crucial for effective transportation management and urban planning. In recent years, deep learning techniques, particularly Long Short-Term Memory (LSTMs) networks, have shown promise in capturing the complex temporal dynamics of traffic systems. This paper proposes a novel approach by integrating LSTM networks with the multi-class METANET model, considering on/off ramps, to enhance traffic state prediction accuracy. We develop both univariate and multivariate versions of the physics-informed multi-class LSTM model, enabling comprehensive exploration of traffic variables such as speed and flow for passenger and heavy vehicles. Experimental results demonstrate the superior predictive capabilities of our approach compared to traditional LSTM models, particularly in capturing the intricate relationship between speed and flow variables. Our findings underscore the significant benefits of incorporating physics-based principles into deep learning architectures for traffic prediction, facilitating more robust and reliable traffic state estimations for real-world applications.

I. INTRODUCTION

Accurate prediction of traffic state is essential for effective transportation management, urban planning, and traffic control systems. Traditionally, traffic state forecasting relied on statistical methods or dynamic models, often struggling to capture the complex temporal dynamics and interdependencies inherent in traffic systems.