The upsurge of mobile service demands underscores the inevitability of the emergence of 6G technology. Amidst the challenges that accompany the 6G networks, the exponential growth of computation-intensive and delay-sensitive applications in vehicular networks poses a significant hurdle. Vehicular edge computing relying solely on terrestrial infrastructures cannot meet the performance requirements due to insufficient resources in edge servers. To address this challenge, we introduce a novel three-layer hierarchical architecture called the air-ground-end aerial computing (AGEAC) architecture, which consists of three types of resource providers, namely high-altitude platform stations, roadside units, and volunteer vehicles willing to share their underutilized resources. We select service vehicles based on the utilization rate of their underutilized resources to ensure fairness. Our objective is to minimize the average latency of all tasks in the long term, and it is formulated as a mixed-integer nonlinear programming problem for task offloading and resource allocation. The complicated intertwined problem is decomposed into four sub-problems by leveraging Lyapunov optimization method. A multi-agent reinforcement learning-based algorithm is designed to make offloading decisions, and convex optimization is introduced to allocate resources iteratively. Extensive simulations have been conducted to demonstrate the superiority of AGEAC in terms of convergence speed, average latency, and resource utilization.

This study proposes an innovative method for high-precision traffic flow prediction through the Fusion of Multi-source heterogeneous External data—FME-GCN. By fusing various external factors, including weather conditions, temperature variations, POI, holidays, and traffic peak and valley periods, FME-GCN significantly enhances the model's perception capabilities. Additionally, by using a lookup function, FME-GCN increases the reference weight of traffic flow data with stronger correlations and improves the model's graph attributes by fusing the node adjacency matrix with POI attributes. Our experiments on the SZ-taxi, PEMS04 and PEMS08 datasets demonstrate that FME-GCN achieves a 1% to 8.1% improvement in prediction accuracy while maintaining a minimal increase in computational complexity (less than 1%). Furthermore, ablation studies further validate the significant contribution of the integrated external factors to the model's performance.

This paper proposes the DMS‐SC (dynamic multi‐semantic sensing and safe control) framework to address the challenge of balancing dynamic optimisation and operational safety in the integration of prediction and control for traffic signal systems. The framework integrates a multi‐semantic spatiotemporal prediction module at the front end, which adaptively integrates three types of semantic graphs‐spatial topology, temporal periodic patterns and dynamic real‐time interactions‐to accurately forecast traffic states and quantify their uncertainties. At the back end, a safety control module based on neural differential equations is employed to construct time‐varying safety sets and a real‐time safety filter, enabling online correction of reinforcement learning policies and theoretically ensuring that all control commands comply with safety constraints. Experiments conducted on multiple real‐world datasets demonstrate that DMS‐SC not only improves prediction accuracy but also outperforms existing methods in reducing delays and enhancing traffic efficiency, while no safety violations were observed during test validation. This provides a reliable and efficient adaptive solution for urban traffic signal control.

In traffic flow prediction, many external factors, such as weather and holiday factors, have a significant impact on traffic flow. However, previous research on traffic flow prediction mainly focused on the internal correlations within traffic flow data, failing to integrate or comprehensively integrate the external factors affecting traffic flow to build prediction models. Based on this, this study collected as comprehensively as possible the data of external factors affecting traffic flow and integrated these data to construct a traffic flow prediction model. This study proposes a Temporal Graph Convolutional Network algorithm (IM - GCN), which comprehensively integrates multi - source heterogeneous external factors that affect traffic flow and for which collectable data are available, for traffic flow prediction modeling. This algorithm expands the receptive field of traffic flow features and enhances the reference weight of data with stronger correlations. The effectiveness of the proposed algorithm was verified through a large number of experiments on widely used traffic flow datasets. Experiments show that, compared with the state - of - the - art algorithms, on the premise that the increase in complexity reflected by the computational time consumption is less than 1%, the accuracy of the prediction model has been improved by 1% - 4%. Ablation experiments experimentally prove that the external factors affecting traffic flow added in this study can effectively improve the performance of the traffic flow prediction model.