Distributed multi-agent reinforcement learning for multi-objective optimal dispatch of microgrids.

Area of Science:

• Power Systems Engineering
• Artificial Intelligence
• Control Theory

Background:

• Distributed microgrids are crucial for reliable and economic power system operation.
• Coordinated economic and environmental objectives are vital for microgrids.
• Existing methods face challenges in handling continuous state changes and resource constraints.

Purpose of the Study:

• To develop a distributed multi-agent reinforcement learning (MARL) algorithm for microgrid optimal dispatch.
• To address multi-objective optimization problems with continuous state and power values.
• To reduce computational and communication resource requirements while ensuring agent privacy.

Main Methods:

• A distributed multi-agent reinforcement learning (MARL) algorithm utilizing an actor-critic architecture.
• Learning multiple critics for subtasks and using neighbor-only information for dispatch strategy.
• Employing linear function approximation to guarantee algorithm convergence.

Main Results:

• The proposed algorithm effectively handles multi-objective optimal dispatch for microgrids.
• Significant reduction in computation and communication resource demands was achieved.
• Agent privacy was successfully protected during information interaction.

Conclusions:

• The developed MARL algorithm demonstrates effectiveness in achieving multi-objective optimal dispatch in microgrids.
• The approach offers a scalable and privacy-preserving solution for distributed microgrid management.
• Simulation results validate the algorithm's performance and practical applicability.