부산대학교 도서관

With the increasing penetration of renewable energy sources in microgrids, optimizing energy management becomes more complex. Classical online deep reinforcement learning (DRL) algorithms often initiate actions in an unsafe manner, potentially leading to equipment damage or resource waste. To enhance safety and feasibility in microgrid scheduling, this study proposes an Offline-to-Online with Expert Data (OOED) DRL, which initializes effective reinforcement learning (RL) control policies by imitating excellent expert behavior without initial interaction, then refines them through online interaction, thereby enhancing efficiency. Experimental results show that OOED outperforms the offline or even online RL by delivering superior solutions in dynamic microgrids environments.