A graph representation learning based on optical network computing power scheduling method is proposed to address the problem of resource scheduling mismatch caused by the independent management of existing computing power resources and optical networks. This method constructs node and edge feature maps and uses graph convolutional networks for clustering, forming a bipartite graph to map the source nodes of computing power services to the destination nodes. By introducing learning factors to optimize the mapping relationship in the bipartite graph, with the goal of minimizing latency, efficient computing power scheduling can be achieved. The simulation results show that the proposed method significantly reduces the blocking rate and average processing delay of services in a multi granularity computing power coexistence environment, and improves resource utilization.