Abstract: The deep reinforcement learning algorithm is data-driven and does not rely on specific models, which can effectively address the complexity issues in virtual power plant (VPP) operation. However, existing algorithms are difficult to strictly enforce operational constraints, which limits their application in practical systems. To overcome this problem, an improved deep Q-network (MDQN) algorithm based on deep reinforcement learning is proposed. This algorithm expresses deep neural networks as mixed integer programming formulas to ensure strict execution of all operational constraints within the action space, thus ensuring the feasibility of the formulated scheduling in actual operation. In addition, sensitivity analysis is conducted to flexibly adjust hyperparameters, providing greater flexibility for algorithm optimization. Finally, the superior performance of the MDQN algorithm is verified through comparative experiments. An effective solution is provided to address the complexity issues in the operation of VPP.