Sustainable Industrial Energy Management: Multi-Agent Deep Reinforcement Learning with GA and Digital Twins

The paper presents an approach towards optimizing energy management in industrial complexes using Multi-Agent Reinforcement Learning (MARL) and Digital Twin technologies, contributing to SDG 9 and SDG 11. Industrial complexes, characterized by diverse energy infrastructures including distributed energy resources (DERs), backup generators, combined heat and power systems, and energy storage units, face significant challenges in ensuring reliable and efficient energy supply while minimizing costs and emissions. To address these challenges, we employ the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm that enables decentralized decision-making for energy management. The Digital Twin of the industrial complex serves as a real-time simulation model, capturing dynamic changes in load demand, renewable generation levels, storage states, market prices, weather forecasts, and grid parameters. The proposed system utilizes a well-defined state space comprising this data to optimize the action space, which includes adjusting loads, scheduling generation units, controlling storage systems, and facilitating energy trading. A Unit Commitment Agent is developed to ensure reliable supply by effectively scheduling generation units. Furthermore, we propose strategies for effective economic dispatch aiming to reduce operational costs and enhance energy efficiency. The results demonstrate the potential of integrating RL and Digital Twin technologies to create smart, adaptive energy management systems that align with the goals of advancing industrial innovation, infrastructure, and sustainable urban development.