Currently, the problem of Quality-of-service in service composition has emerged as a critical research topic in the fields of cloud computing and service computing. Aggregating multiple existing services based on QoS awareness to meet complex functional and non-functional requirements, which simple services cannot satisfy, and choosing services for aggregation in a rapidly changing service environment to achieve the highest possible service quality, poses an NP-hard challenge. Traditional heuristic algorithms, while efficient to an extent, often show limitations in adapting to this rapidly changing environment as they typically rely on static rules or preset conditions, lacking the ability to learn from and adapt to environmental changes. Methods based on reinforcement learning are the most commonly used approach in dynamic and complex environments. However, as the scale of composition increases, they face challenges such as constantly changing service characteristics, user demands, and a vast state space. Therefore, this paper designs a service composition framework based on graph convolution and Deep Reinforcement Learning (DRL), which extracts environmental states from three dimensions: abstract services, specific services, and QoS features, using graph convolution, makes decisions using DRL, and reduces the state space with a double channel action network. Simulations have demonstrated that our method surpasses existing methods in terms of service composition quality, scalability, and stability.