Deep reinforcement learning-based hierarchical voltage control considering partition dynamic aggregation

• A hierarchical voltage control method for flexible distribution networks based on dynamic zonal aggregation is proposed. • Node electrical distances are evaluated using power–voltage sensitivity and adjustable capacity of distributed energy resources. • An attention mechanism is employed to prioritize the reduction of voltage deviations. • Multiple zonal agents are designed to allocate weights for multi-objective control. To address the voltage stability challenges in flexible distribution networks (FDNs) caused by the high penetration of distributed energy resources (DERs), this paper proposes a deep reinforcement learning (DRL)-based hierarchical voltage control strategy considering partition dynamic aggregation. The proposed strategy partitions lower-level agents based on the aggregation results of DERs to achieve voltage optimization within each aggregated region, while an upper-level agent coordinates the allocation of multi-objective weights, prioritizing the mitigation of overall voltage deviations in the FDN. Specifically, a dynamic aggregation partitioning algorithm is developed within the lower-level control structure, where electrical distance is quantified based on power–voltage sensitivity and available capacity, enabling accurate partitioning of DERs according to electrical proximity. Subsequently, aiming to prioritize voltage deviation governance within a multi-objective control, a multi-objective attention-based deep deterministic policy gradient (MOADDPG) algorithm is designed. The algorithm incorporates an attention mechanism to dynamically weight individual objective functions and employs multiple critic networks to allocate weights across objectives, thereby prioritizing the mitigation of voltage deviations. Finally, the upper-level agent reduces global voltage deviations by coordinating multi-objective weight allocation based on the regulation outcomes of the lower-level agents. Simulation results on the 65-bus system demonstrate that, compared with the optimal performance of algorithms such as deep deterministic policy gradient (DDPG), the proposed MOADDPG algorithm achieves the lowest steady-state voltage deviation of 0.0186p.u. The average voltage deviation is reduced by 7.0% and 2.6% compared with the proximal policy optimization (PPO) and DDPG algorithms, respectively. In addition, compared with the uncontrolled scenario, the proposed method reduces the peak voltage deviation by 4.86%. These results indicate that the proposed strategy effectively mitigates voltage deviation issues caused by the fast dynamic characteristics of the distribution network.