Hierarchical graph reinforcement learning for coordinated dispatch and reconfiguration of active distribution networks

Traditional studies on active distribution networks often optimize operational dispatch and network topology reconfiguration in a decoupled manner, although the two decisions are strongly coupled through power-flow redistribution, controllable-device outputs, and network constraints. This paper develops a problem-specific hierarchical coordination framework in which established DDPG, DDQN, and GAT components are used as implementation tools rather than claimed as standalone algorithmic innovations. The dispatch layer first determines continuous device actions. An AC power-flow update then produces the post-dispatch operating state, which is encoded together with the current topology by a two-layer multi-head GAT and transferred to the reconfiguration layer. The DDQN reconfiguration agent subsequently selects only from pre-screened radial and AC-feasible topology actions. Studies on modified IEEE 33-, 69-, and 123-bus benchmark systems show lower operating cost, network loss, and voltage deviation than the comparison methods. The 24 h IEEE 33-bus switching log contains three branch exchanges, corresponding to six switch operations, which confirms that the obtained topology policy does not rely on frequent switching. The larger benchmark cases further provide cross-system evidence of computational efficiency; however, these standard systems are not used to claim utility-scale scalability or guaranteed real-network online applicability.