Mobile Edge Computing (MEC) enables flexible resource allocation but faces challenges in scheduling complex workflows represented as directed acyclic graphs (DAGs). Existing methods often fail to address multi-objective conflicts in multi-workflow scheduling, especially in task sequencing and resource mapping. To solve this, we propose the Graph-Temporal Double Deep Q Network (GT-DQNet). This approach begins by utilizing topological ordering to preprocess the task dependencies, thereby preserving the sequential relationships between tasks. A Neighborhood-Aware Graph Attention Network (NAGAT) is then employed to aggregate spatial features, to capture the structural relationships within edge nodes. To model temporal dependencies in task scheduling, the Multi-Layer Perceptron (MLP) module in the Double Deep Q-Network (DDQN) is replaced with a Long Short-Term Memory (LSTM) network, enhancing decision-making through temporal feature learning. Finally, despite the integration of multiple modules, GT-DQNet maintains low computational overhead. Experimental results on the Alibaba Cluster Trace dataset show that it significantly reduces energy consumption and latency, with its advantages becoming especially prominent under heavy workloads.