Scalable Multi-Agent Model-Free Demand Response for Voltage Regulation in Grid-Interactive Efficient Buildings

This paper proposes a scalable model-free multi-agent deep reinforcement learning (MADRL) framework for voltage regulation in grid-interactive efficient buildings. Unlike traditional methods that rely on reactive power control, the proposed approach utilizes active power adjustment through intelligent demand response (DR) scheduling. The architecture features a decentralized control structure, where customer agents optimize their appliance usage based on dynamic incentives from an aggregator agent. The optimization problem considers various constraints such as user comfort, electricity pricing, voltage deviation penalties, and the presence of distributed photovoltaic (PV) generation. A multi-objective function integrating dynamic price signals, user dissatisfaction, and voltage deviation is formulated. The aggregator leverages voltage-aware incentive signals to nudge consumers toward grid-supportive load behaviors. Simulation investigations are curried out to show that the MADRL framework reduces peak and mean load, improves voltage stability, and preserves user privacy. The paper aims to demonstrate the potential of decentralized, model-free DR systems in modern distribution grids.