A Multi-Agent Deep Reinforcement Learning Framework for Real-Time Energy Management in Smart Building Communities

Abstract

The rising adoption of distributed energy resources and dynamic electricity pricing demands more effective approaches for energy management in smart connected buildings. This paper presents a novel multi-agent deep reinforcement learning (MADRL) framework for coordinated energy scheduling in a community of multiple smart buildings. The proposed approach manages heating, ventilation, and air-conditioning (HVAC), household appliances, and electric vehicle (EV) charging under real-time conditions. Each building is equipped with a MADRL agent that aims to minimize electricity costs and battery degradation while ensuring occupant comfort. The method integrates a multi-agent soft actor-critic (MASAC) algorithm with a shared attention mechanism to output both discrete and continuous actions for different types of assets in smart buildings. It adopts a centralized training and decentralized execution (CTDE) paradigm to preserve privacy and scalability. Simulation results demonstrate that the proposed framework achieves significant cost savings, ranging from 11.89% to 23.17% compared to alternative optimization-based and MADRL controllers, and effectively performs cost reduction and peak shaving. Additionally, the proposed model enables the controllers to make decisions in milliseconds after offline training, which makes it ideal for real-time energy management scenarios. These findings underscore the practicality and efficiency of the proposed method, enabling flexible demand response and sustainable load scheduling in modern building communities.

Keywords: Multi-agent deep reinforcement learning, Energy management system, Connected building community, Demand Response