With the development of technology and society, the traditional energy system has become difficult to meet the demand. Applying Deep Reinforcement Learning (DRL) to solve scheduling problems in microgrid cluster edge-cloud collaborative architecture provides a new solution. However, there is no research work has been completed to address the independent training, deployment, and inference of DRL in the microgrid cluster scenario. In this paper, we propose a federated DRL-based request scheduling algorithm for distributed microgrid cluster scenarios with the goal of maximizing the long-term utility of the system. In addition, we prune the DRL model to make it more applicable to resource-constrained edge nodes. The experimental results show that the proposed algorithm has more stable performance and better adaptability to the dynamic system environment compared to the traditional centralized training. In addition, the pruning of the model compresses the size of the model to 50.4% with a 4% loss of accuracy.

Service scenarios under edge-cloud collaboration are becoming more diverse in terms of service performance requirements. For example, smart grids require both intelligent control and long-term optimization, which poses considerable challenges for service providers to meet quality of service (QoS). However, current pioneering work has not yet explored both system utility and QoS guarantees. Therefore, this paper investigates the optimization problem of edge-cloud collaborative scheduling for QoS guarantees. First, we model the edge-cloud collaborative scheduling scenario and derive two sub-problems such as service deployment and request dispatch. Second, we design a near-optimal scheduling algorithm based on a submodular function optimization approach with the objective of maximizing the number of requests that are processed within the edge-cloud cluster under QoS constraints. Finally, our experiments verify the beneficial effects of the proposed algorithm in terms of throughput rate, scheduling time cost, and resource utilization.