A novel operation method for renewable building by combining distributed DC energy system and deep reinforcement learning

Reducing carbon emissions has been a focus problem with the rapidly increasing building energy consumption. One solution is adopting more Renewable Energy Resources (RESs) for building energy supply. To overcome the intermittence of RESs, researchers paid efforts in flexible demand response based on centralized operation and model-based control, however, which get challenges for scalability and uncertain dynamic building systems. Moreover, few works have considered user willingness as an important part of human-machine interaction and user satisfaction. Thus, we propose a novel operation method called DC-RL for renewable building energy systems. DC-RL designs a distributed DC energy system, which is scalable, control-friendly, and provides users the willingness option for flexible operation. For energy control, DC-RL adopts a model-free deep reinforcement learning (DRL) algorithm Soft-Actor-Critic (SAC) to adjust demand to matching renewable supply with maintaining user satisfaction. We evaluate DC-RL on two real-life datasets. Compared to baselines, DC-RL improves energy saving and PV self-consumption by 38% and user satisfaction by 9%. DC-RL achieves near-zero-carbon buildings with 93% self-sufficiency rate and reduces up to 33% of battery dependency.