Safe Decentralized Operation of EV Virtual Power Plant with Limited Network Visibility via Multi-Agent Reinforcement Learning

A team of researchers has published a paper proposing a novel AI framework called TL-MAPPO (Transformer-assisted Lagrangian Multi-Agent Proximal Policy Optimization) to solve a critical challenge in modern power grids. As electric vehicle (EV) adoption surges, charging stations become major assets for virtual power plants (VPPs) but also pose risks to local grid voltage stability. The core problem is that VPP operators typically have only partial, aggregated visibility into the distribution network's real-time state, making safe and efficient coordination difficult.

The proposed TL-MAPPO framework tackles this by using a multi-agent reinforcement learning (MARL) approach. Each EV charging station (EVCS) operates as a decentralized agent that learns its own charging policy. Crucially, the agents are trained centrally with a safety mechanism called Lagrangian regularization, which enforces hard constraints on voltage limits and charging demand. To improve decision-making with limited data, each agent is equipped with a transformer-based embedding layer that captures complex temporal patterns in electricity prices, loads, and charging demand.

In experiments conducted on a realistic 33-bus power distribution network model, the TL-MAPPO system demonstrated significant performance gains. It reduced voltage limit violations by approximately 45% and lowered overall operational costs by about 10% compared to existing multi-agent deep reinforcement learning baselines. This performance was achieved under the realistic constraint of limited network visibility, where agents must rely on aggregated information rather than full grid state data. The research highlights a practical path toward deploying AI for safer and more economical management of the growing fleet of distributed energy resources, a key requirement for achieving net-zero power systems.