Scalability of MARL for Active Voltage Control and viability of per-inverter agents

Determine whether multi-agent reinforcement learning approaches for active voltage control that partition a distribution network into regions with a single controlling agent per region scale effectively as the number of agents increases, and ascertain whether modeling each inverter as an independent agent (one agent per inverter) can maintain reliable performance and stability in distribution networks.

Background

Prior studies on applying multi-agent reinforcement learning (MARL) to active voltage control often divide the distribution network into several regions, assigning a single agent to control each region. While such setups have shown promise, they raise questions about how well these methods scale as the system size and number of agents grow.

The authors highlight uncertainty about whether these regional MARL schemes retain performance with more agents and whether a more granular design—treating each inverter as an independent agent—can be effective. Addressing these questions informs the design of MARL architectures for real-world distribution networks with potentially large numbers of controllable devices.