With the rapid development of network technologies such as artificial intelligence and the Internet of Things, the Internet of Vehicles, as an emerging information and communication technology, is widely used in the field of intelligent transportation systems. As more attention is paid to data privacy, an increasing number of vehicles are reluctant to share local data. To address this issue, federated learning has been introduced to the Internet of Vehicles as an emerging learning paradigm, allowing models to be trained without directly accessing raw local data. However, frequent communication between vehicles and roadside unit nodes can lead to a decrease in the accuracy of the global model in federated learning, as well as slower model convergence and higher communication overhead during training. Therefore, this paper proposes a node-based dynamic communication scheme based on federated deep reinforcement learning. A global model accuracy optimization objective is constructed based on real-time network conditions, vehicle communication capabilities, and task requirements. To address the vehicle node selection problem, an adaptive algorithm is designed to select the optimal set of vehicles for global model aggregation. Then, a dynamic asynchronous aggregation strategy is used to improve the efficiency of model training. By comparing with other baseline algorithms on standard datasets, the proposed method achieves an average accuracy improvement of 3% compared to traditional methods, while reducing communication latency and decreasing energy consumption by an average of 37%.