Consumer-Centric Home Energy Management System Using Trust Region Policy Optimization-Based Multi-Agent Deep Reinforcement Learning

Autonomous home energy management system (HEMS) is the key to improving energy efficiency in the active distribution network. This HEMS also needs to maintain customer satisfaction while maximizing cost savings under dynamic price conditions, incorporating uncertainties of consumer behavior, and renewable energy generation. In this paper, a consumer-centric HEMS using Trust Region Policy Optimization (TRPO) based multi-agent deep reinforcement learning (DRL) is presented. This Multi-Agent TRPO (MA-TRPO) based HEMS is trained to respond to the dynamic retail price and the local energy generation by scheduling the Interruptible-Deferrable load (IDA) and Battery Energy Storage System (BESS). Five-minute retail electricity price derived from wholesale market price and the PV generation data derived from real-world PV profiles are used to train the proposed MA-TRPO-based HEMS in discrete action space. The performance of the proposed HEMS is relatively better than the existing policy-gradient-based on-policy approaches such as Proximal Policy Optimization and Policy Gradient-based HEMS as validated via training and testing using the same dataset.