Accurate traffic forecasting plays an indispensable role in modern Intelligent Transportation Systems (ITS). Recent years have witnessed significant advancements in the modeling capabilities and inference efficiency of continuous-time generative models, which learn a dynamic system to transform a simple prior distribution into the target data distribution. Motivated by the emerging generative paradigm, we rethink the traffic forecasting task from the perspective of flow matching. Specifically, we propose to learn a velocity field that transports samples from a prior distribution to the distribution of future traffic values. We introduce two flow matching models based on different assumptions on prior distribution: the commonly used standard Gaussian distribution and the distribution of the historical traffic values. Furthermore, our approach incorporates the latest mean flow techniques to enable one-step generation, ensuring simplicity and high efficiency in both training and inference phases. In experiment on real world data, our proposed models demonstrate desirable performance.

To implement proactive traffic management, traffic forecasting becomes an essential function of modern intelligent transport systems (ITS). Traffic flows on motorways exhibit substantial variability, making it necessary to capture high-frequency patterns in the spatiotemporal model. To address the challenges, a representation learning approach is leveraged in this paper to extract high-level features that facilitate traffic forecasting on motorway. A bottom-up learning structure is proposed to sequentially extract information from local to the global level. Computational experiments show that simple models with informative representation may achieve satisfactory performance for traffic prediction.

Monitoring and forecasting of road traffic conditions is a common practice for real traffic information system, and is of vital importance to traffic management and control. While dynamic traffic patterns can be intuitively represented by space-time diagrams, this study proposes a new concept of space-time image (ST-image) to incorporate physical meanings of traffic state variables. We therefore transform the forecasting problem for time-series traffic states into a conditional image generation problem. We explore the inherent properties of the ST images from the perspectives of physical meaning and traffic dynamics. An innovative deep learning based architecture is designed to process the ST-image, and a diffusion model is trained to obtain traffic forecasts by generating the future ST-images based on the historical patterns.