Dynamic graph learning is now a significant area of study in the field of graph learning because to the quick expansion of social media, traffic networks, and financial monitoring. Traditional GNNs (GNNs) assume that the graph structure remains unchanged during the learning process, which makes them face performance bottlenecks when processing dynamic graphs. To address this problem, this study proposes a reinforcement learning enhanced graph neural network (RL-GNN) model, which aims to improve the performance of GNNs in dynamic graphs through the policy optimization mechanism of reinforcement learning. The RL-GNN model combines the node representation learning of GNNs with the policy optimization of reinforcement learning, and dynamically adjusts the training strategy of GNNs to adapt to the dynamic graph environment. Through 100 rounds of experiments on multiple real-world datasets (including social networks, traffic networks, and biological networks), the results show that RL-GNN improves the accuracy by 8% to 12% and the F1 value by 9% to 14% compared with traditional GNN models. In addition, RL-GNN shows higher stability and efficiency than traditional methods in handling node feature aggregation and adaptability to graph structure changes. Experiments show that the model effectively improves the performance of GNNs in dynamic graph learning, provides a new solution, and promotes the application of GNNs in complex dynamic graph tasks.