Answers are grounded in this paper and cite their sources.
The high penetration of residential distributed energy resources (RDERs) in distribution networks exhibits high randomness and spatiotemporal coupling, which may cause voltage fluctuations. Although multi-agent deep reinforcement learning (MADRL) enables collaborative control, its reliance on extensive real-world interaction data increases training costs and operational risks. To address these challenges, a collaborative voltage regulation strategy, termed DWM-MATD3, is developed by integrating MADRL with a dual-layer world model (DWM). The upper layer models distribution-network power-flow dynamics using a heterogeneous graph attention network and gated recurrent units (HGAT-GRU), thereby capturing network topology, temporal dependencies, and physical power-flow constraints. The lower layer models residential-device behavior using MLP-GRU networks to improve virtual-environment fidelity. A three-stage training scheme comprising real-data pretraining, real-virtual interaction fine-tuning, and hybrid-data training reduces dependence on real-world data. Parameter sharing further narrows the simulation-to-reality gap and improves sample efficiency. Experiments on modified IEEE 33-bus, 141-bus, and 322-bus systems demonstrate that DWM-MATD3 consistently outperforms the baselines in terms of voltage-violation rate and line loss, with its performance gains increasing as the system scale expands. The testing results demonstrate that the presented DWM-MATD3 method significantly improves the voltage regulation performance in distribution networks with residential distributed energy resources.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: