<p>Timely collection and analysis of large-scale vehicle travel data are crucial for traffic flow prediction. However, centralized data processing poses significant privacy risks, while federated learning within dynamic Internet of Vehicles (IoV) faces challenges such as device heterogeneity, non-independent and identically distributed (non-IID) data and stochastic communication latency issues. To address these issues, we propose a semi-asynchronous federated learning method, FedSAP. Firstly, FedSAP constructs a global model by aggregating model parameters from any minority of clients. This approach mitigates the negative impacts of unstable and unreliable communication in the IoV on model training. Secondly, a dynamic asynchronous aggregation strategy is implemented at the server side. This strategy dynamically adjusts global model aggregation weights based on the staleness of local models and differences in gradient directions between local and global models, addressing performance degradation caused by staleness. Additionally, at the vehicle terminal, FedSAP introduces a parameter decoupling technique. By dynamically adjusting the personalization coefficient, the global model’s adaptability to local sensing data is accelerated, reducing the negative impact of data heterogeneity on its performance. Experiments on public traffic flow datasets demonstrate that FedSAP outperforms baseline methods on two common traffic flow prediction models, achieving approximately a 50% increase in convergence efficiency and a 9% reduction in prediction error. The results substantiate the superiority and adaptability of the proposed method in dynamic vehicular networking environments.</p>