Integrating Fusion Autoencoder with Multi-Agent Reinforcement Learning for Optimal Energy Dispatch Under Uncertainties

Energy hub concept has been applied to energy management systems to enhance system flexibility. However, energy management studies face various challenges due to demand uncertainties, system complexity, and interactions between different energy carriers in energy hubs. This study considers smart operational strategies for a novel green energy hub (GEH) to supply power and heat to a commercial community. To handle demand uncertainty, a fusion autoencoder (FAE) model was used to reconstruct any abnormalities in demand data. To enhance flexibility, the GEH is equipped with storage facilities, including a battery bank and thermal storage. The scheduling of the GEH was optimized using a multi-agent reinforcement learning (MARL) model. This approach considers a balance between centralized training and decentralized dispatch to minimize the total cost of the GEH system while meeting demand. The FAE accurately reconstructs abnormal demand patterns, achieving high accuracy with
$\mathbf{R}^{\mathbf{2}}$
values of 0.9806 and 0.9407 for electric and thermal load abnormalities, respectively. By combining advanced FAE with MARL-based optimization, optimal dispatch strategies are developed, enabling decentralized policies that meet demand while minimizing resource losses. Comparative analysis shows substantial cost savings, with daily costs reduced from $2737.05 USD to $745.59 USD by transitioning from abnormal to reconstructed data, saving $1991.46 USD per day. Thus, integrating FAE with MARL proves pivotal in optimizing system performance and resource utilization in green energy systems.