<p>With the wide access of photovoltaic (PV) units in various scenarios, the optimal scheduling problem of distributed PV microgrids (MGs) has become increasingly complex and important. However, existing model-based and model-free methods suffer from the problems of inefficient handling of nonlinear effects and constraint overflow when dealing with uncertainties and complex operational constraints. Deep reinforcement learning (DRL), as a model-free method, can learn the dynamic properties of the system by interacting with the environment, but it is prone to violating the operational constraints, leading to a decrease in the reliability of the decision. In addition, training DRL requires a large amount of MG operational data, and privacy and security issues need to be addressed. These issues are addressed in this study by proposing a strongly constraint-aware two-stage optimization algorithm named the FSC–DQN algorithm. The method combines the discrete action space processing capability of the DQN and the feature expression capability of deep neural networks (DNNs); it also rationally models this DNN as a mixed integer programming (MIP) problem with the goal of minimizing constraint overflow as the reward objective to achieve strong constraint awareness of complex operations. Moreover, a federated learning mechanism is introduced to achieve the optimized update of distributed parameters of the MG system and to ensure data privacy and security through the framework. The results show that the algorithm can effectively optimize the operation scheduling of MGs while satisfying the operation constraints and achieve robust performance in various scenarios.</p>