The large-scale integration of distributed energy resources and diverse regulating devices makes fast, optimal voltage control in distribution networks increasingly challenging. Existing multi-agent deep reinforcement learning methods under the centralized training and decentralized execution framework suffer from poor scalability in large-scale systems due to the exponential growth of joint action spaces, while centralized training and centralized execution achieve better control performance but rely heavily on full-network communication. To address these issues, we propose a dual-timescale voltage control strategy based on key observations and a minimal control set. First, hierarchical clustering using electrical distances partitions the network into weakly coupled regions. An attention mechanism then evaluates node importance to select key observation nodes that represent regional voltage states, compressing the state space. Second, voltage sensitivity analysis with attention weight correction ranks controllable resources by voltage regulation contribution to construct a minimal control resource set. Finally, a case study on the IEEE 123-bus system shows that the proposed method reduces communication and control scale while achieving effective voltage regulation and loss reduction.
Haotian Su (Sun,) studied this question.