The multi-energy coupling characteristics and increasingly complex topology of integrated energy systems (IES) make optimal scheduling a pivotal challenge in balancing economy efficiency and operational security. To address the issues of convergence difficulty caused by the curse of dimensionality in traditional multi-agent reinforcement learning, as well as local optima resulting from insufficient exploration mechanisms, a real-time optimal scheduling method based on a prior knowledge-guided multi-stage QMIX architecture is proposed. First, the IES real-time optimal scheduling is formulated as a distributed partially observable Markov decision process, and a QMIX framework based on joint action value function updates is constructed. Then, according to the coupling relationships among energy units, a clustering-based multi-stage QMIX training strategy is designed to alleviate the curse of dimensionality. Finally, an enhanced action exploration mechanism incorporating prior knowledge is developed to guide the convergence trajectory. Scheduling simulations are conducted under multiple load scenarios (40 sample days). The results show that the proposed method exhibits significant advantages in convergence performance and effectively reduces the overall system operation costs.