Deep reinforcement learning control architectures for industrial multi-energy systems: from single-agent to hierarchical multi-agents

The increasing share of renewable energy sources in industrial, multi-energy systems has introduced significant challenges for the optimal coordination of multiple energy carriers. This work focuses on improving the cost-effective and robust operation of an industrial plant that simultaneously produces electricity, steam, hot water, and chilled water. It explores whether data-driven control strategies based on deep reinforcement learning can enhance the economic and energy performance of the plant, compared to traditional methods, and whether multi-agent structures can further improve robustness. Three control architectures were designed and evaluated – centralized, decentralized, and hierarchical – using a year of actual operational data from the industrial plant. The results show that all of the proposed strategies reduced total operating costs by more than 6% compared to existing rule-based control. The hierarchical configuration achieved the best performance and demonstrated superior robustness to variations in energy prices. These findings highlight the potential of learning-based hierarchical coordination as a practical and resilient framework to manage complex industrial energy systems.