Scheduling workflows with complex dependencies under dynamic resource availability in heterogeneous clouds remains highly challenging. Static and heuristic methods often fail to adapt to workload and resource changes, resulting in higher makespan, lower utilization, and increased cost. We propose GDST-PPO, a graph-based dynamic scheduling framework that encodes workflow DAGs with a GNN, models temporal and resource contexts via a Transformer with sparse attention, and optimizes decisions using PPO with a multi-objective reward. We perform extensive experiments on five benchmark workflows under realistic heterogeneous cloud settings. Across six baselines, GDST-PPO achieves up to 10.3% lower makespan, 18.1% higher resource utilization, and an 11.4% gain in overall score. In absolute terms, it attains a makespan of 7,168,820s, resource utilization of 39.55%, and an overall score of 6.62 on our benchmark, demonstrating efficient, flexible, and cost-effective cloud workflow management.