A Cloud-Edge Intelligence-Based Optimization Method for Distribution Network Partitioning and Operation Considering Simulation Inaccuracy

Abstract

The increasing emergence of distributed renewable generation and varying load demand adversely affect the security of distribution network operation. In this paper, a cloud-edge intelligence-based optimization method is proposed for distribution network partitioning and operation to derive the near-optimal real-time control strategies of switches, energy storage systems, static var compensators, and capacitor banks. It realizes centralized training in the cloud and real-time execution at edge. To address the computational burden in large-scale distribution networks, a novel partitioning method is devised to facilitate network division for operation optimization. Then, a new switch importance calculation approach is introduced to reduce the dimensionality of switch action space. Next, a multi-agent Markov Decision Process is established, where each agent corresponds to a type of controlled devices in each sub area. Finally, considering the specific inaccuracies in the distribution network model, a modified domain randomization method and an improved mixed multi-agent soft Actor-Critic algorithm is developed to enhance the robustness of policies under mismatch between the simulation model and the practical system. Numerical studies in IEEE 33-bus system and a practical 445-node distribution network are implemented to validate the effectiveness and merits of the proposed optimization method.