부산대학교 도서관

Deep Reinforcement Learning (DRL), with its data-driven nature and model-free advantage, has attracted great interest in the field of microgrid energy management system. The choice of reward mechanism plays a crucial role in the performance and effectiveness of DRL-based microgrid energy management. This paper aims to investigate the reward mechanism design by comparing the performances of DRL-based microgrid energy management under two different reward mechanisms, namely, cliff walking pattern and Leduc poker pattern. The reward mechanism incorporates auxiliary rewards alongside the primary reward to harmonize diverse objectives. Using a real microgrid dataset, the performance of DRL agents under different reward mechanisms are compared. The experimental results demonstrate that different reward mechanisms have a significant impact on the convergence speed and generalization ability of trained microgrid energy management policy.