Optimal Energy Scheduling for Communities in Smart Microgrids with Multi-Agent Reinforcement Learning

作者：Hong Yang, Xun Shao, Go Hasegawa, Hiroshi Masui · 发表于：IIAI International Conference on Advanced Applied Informatics · 年份：2023 · DOI：10.1109/iiai-aai59060.2023.00081 · 研究领域：Computer Science

Energy management in buildings is becoming increasingly important with the proposed carbon neutrality target. Home energy management systems(HEMS) can reduce electricity costs and carbon emissions by scheduling the energy consumption of household loads and using solar power to generate electricity. Cooperation between multiple HEMS can more effectively reduce electricity costs under the peak load limit. To minimize the electricity cost of multiple users under the peak pricing scheme, we first formulate the total cost minimization problem as a Markov game. Then we propose a multi-agent deep deterministic policy gradient (MADDPG)-based energy consumption scheduling algorithm for multiple household loads to address the partial observability of the environment and the instability of the environment from the agent’s perspective. The algorithm adopts a centralized learning but decentralized execution approach to achieve cooperation among agents to meet peak power constraints. Simulation results show that our online deep reinforcement learning method can reduce the peak-to-average ratio of total energy consumption and electricity cost for all households based on observational information and electricity prices.