In the edge computing scenario, computing nodes are composed of various heterogeneous devices with significant differences in performance and resources. Moreover, the mobility of edge devices results in frequent changes in network topology. The time dependencies between tasks further increase the difficulty of efficient scheduling and network load balancing. This paper proposes an innovative framework combining task flow Partitioning and dynamic task scheduling. First, the task flow is divided into sub graphs as independent scheduling units to optimize the task execution order, and then tasks are scheduled based on task requirements, task dependencies, and real-time network status. Additionally, we introduce a feature extraction method using a GCN based on a location-aware attention mechanism, which extracts node features from the physical network topology graph, and enhances these features. We also propose an average incremental A3C architecture to promote stable convergence of the neural network and avoid conflicts in scheduling decisions among multiple agents. We conduct extensive simulation experiments based on Alibaba’s real dataset. The results show that the proposed algorithm significantly outperforms existing baseline algorithms in key performance indicators such as total task processing delay, load balancing degree, and task success rate.