Multi-agent deep reinforcement learning for Smart building energy management with chance constraints

With the development of building power supply technology, traditional building is gradually replaced by smart building (SB) with advanced energy management system, enabling effective management of building energy resources. However, the uncertainty of photovoltaic (PV) output brings new challenges to building energy management. Therefore, this paper proposes a multi-agent deep reinforcement learning-based energy management strategy for SB, in which SB is decomposed into multiple energy-local area networks (E-LANs) with controllable devices, each E-LAN is then regarded as an agent. According to the multi-agent deep deterministic policy gradient algorithm, each agent learns the optimal energy management strategy for E-LAN through interactions with the environment, thereby achieving overall energy management for SB. To fully account for the uncertainty of PV outputs, first, random PV output time sequences are used during training process of algorithm. Then, the equivalent PV output is obtained according to the converted deterministic constraints from the joint chance constraint of the original problem, and is used for solving the day-ahead energy management. Simulation results show that compared to stochastic programming-based method and deep deterministic policy gradient algorithm-based method, the proposed energy management method reduces the total cost by up to 11.5% within a scheduling period and by up to 7.6% in 3 continuous scheduling period. Additionally, energy interaction between E-LANs is improved significantly to promote local energy consumption.