As renewable energy sources such as solar and wind are increasingly integrated into the grid at high proportions

the optimization and scheduling of distributed power generation face challenges due to frequent changes in system topology

affecting the stability and economic operation of the distribution network. Existing methods

designed for systems with fixed topology

rely on precise models and are time-consuming

making real-time control difficult. Current deep reinforcement learning approaches struggle to balance distributed training and mixed discrete-continuous action spaces. This study introduces a distributed power optimization strategy based on multi-agent deep reinforcement learning with two stages and parameter sharing. Initially

the problem is vertically decoupled

constructing a dynamic distribution network reconfiguration model with distributed generation using mixed-integer second-order cone programming to determine the topology. Subsequently

the distribution network environment is horizontally decoupled into several regions. In the second stage

a centralized training with decentralized execution framework that incorporates parameter sharing is proposed. This framework incorporates a multi-agent prioritized double-delay deep deterministic policy gradient algorithm with a priority experience replay mechanism. Topology information is embedded into the distribution network environment

mapped to agents through power flow calculations to minimize network active power loss in the optimization scheduling model. Case studies demonstrate that the proposed algorithm

by considering changes in the distribution network topology and enhancing learning efficiency through strategy and experience sharing among agents

as well as priority experience replay

meets the efficiency requirements of real-time online decision-making and shows superior voltage stability and loss reduction performance compared to other strategies.