In fault-tolerant quantum computing systems, T gates consume more fault-tolerant resources. In this paper, we propose a T-count reduction method based on the Proximal Policy Optimization (PPO) algorithm, minimizing the number of T gates in quantum computations. Initially, within the framework of ZX-calculus graphical language, quantum circuits are transformed into ZX-diagrams. Subsequently, the PPO algorithm is employed to learn a policy that predicts optimal transformation trajectories. To effectively leverage the topological structure of ZX-diagrams, we employ graph neural networks (GNNs) to encode the policy trained via PPO algorithm, while identifying possible transformations through the local structural properties of individual nodes or edges. The proposed method achieves an average 10.17% reduction in T-count under optimal conditions, demonstrating its capability in reducing the number of T gates.