50% seminar: Safe Deep Reinforcement Learning Based Energy Management for Educational Buildings With Guaranteed Constraints

Driven by sustainability goals and the increasing integration of renewable energy sources, effective energy management systems are crucial for enabling smart buildings to operate efficiently. This research proposes a safe real-time energy management system based on safe deep reinforcement learning. The deep deterministic policy gradient (DDPG) algorithm is utilized to handle uncertainties in solar PV generation and energy demand without explicit forecasting. Projection layers are embedded into the DDPG algorithm after the actor and target actor networks to ensure that all actions remain within operational constraints. A case study conducted on an educational building in Stockholm demonstrates the effectiveness of the proposed method in maintaining user comfort, satisfying EV charging requirements, ensuring dishwasher operation, and effectively utilizing the flexibility of both EVs and the dishwasher.

Main supervisor: Qianwen Xu