Abstract: To address slow model convergence， poor training stability， and insufficient data coverage caused by client heterogeneity in vehicular-to-everything （V2X） federated learning， a dynamic weighting and clustering-constrained client selection method based on the soft actor-critic （SAC） algorithm is proposed. First， a vehicular federated learning system model characterizes key performance factors by integrating communication cost， computational resources， and data distribution heterogeneity， formulating client selection process as a multi-objective dynamic optimization problem. Subsequently， the SAC algorithm dynamically adjusts the importance weights of data quality， low latency， and historical contribution according to the current system state， enabling adaptive client selection strategy optimization. Meanwhile， a clustering-constrained mechanism prevents long-term selection concentration on similar clients. By clustering heterogeneous client features， the proposed method enhances the data coverage and diversity of the selected client set. Finally， an asynchronous aggregation mechanism improves training efficiency and reduces latency. Numerical simulation results demonstrate that， compared with random selection， FedCS， and HiCS-FL methods， the proposed method accelerates convergence， improves average accuracy by up to 2.93%， and increases data quality by 19.92%， 23.06%， and 36.03%， respectively. It also reduces global data distribution discrepancies by 13.59%， 17.86%， and 21.85%， respectively. Consequently， this method enhances client data representativeness， expands overall data coverage， mitigates model oscillation caused by non-IID data distributions， and improves the stability and generalization performance of federated learning models in V2X scenarios.