With the dual-carbon goals, the intensified coupling of heterogeneous energy sources is driving integrated energy system (IES) topology to evolve towards greater complexity and flexibility. However, the existing optimal scheduling methods do not sufficiently consider the knowledge of non-European network topology and its heterogeneous power flow constraints. To address this issue, this paper proposes an optimal dispatch method based on graph reinforcement learning. First, by guaranteeing diverse node states, the heterogeneous energy network topology connections are converted into network graph models using graph theory. Second, a state-action-reward framework based on real graph mapping is established, and the graph reinforcement learning method is employed to learn the non-Euclidean knowledge and heterogeneous flow constraints brought by the graph model, thus achieving safe and optimal scheduling of IES. Finally, the real data of an industrial park is used for simulation verification. Compared with the traditional method, the proposed method effectively alleviates the node voltage over-limit problem. The results indicate that the proposed method can achieve optimal dispatch of IES while considering the actual topology operation state information and heterogeneous flow safety.