Optimal control in microgrid using multi-agent reinforcement learning.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Control Systems

Background:

• Microgrids require sophisticated control strategies for cost optimization and stable operation, especially when connected to the main grid.
• Existing methods often struggle with the complexity and dynamic nature of microgrid energy management.

Purpose of the Study:

• To develop an improved reinforcement learning (RL) method for minimizing electricity costs in grid-connected microgrids.
• To ensure power balance and adhere to generation limits of individual units within the microgrid.

Main Methods:

• Formulation of an objective function for optimal microgrid control.
• Introduction of an 'Average Electricity Price Trend' state variable to simplify microgrid dynamics.
• Development of a multi-agent architecture with defined agents, state, action variables, and reward functions.
• Implementation of dynamic hierarchical reinforcement learning based on key state variable change rates.

Main Results:

• The proposed RL method effectively addresses the 'curse of dimensionality' in complex systems.
• Faster learning and policy exploration are achieved in large-scale, unknown environments.
• Simulation results validate the method's effectiveness in optimal microgrid control.

Conclusions:

• The presented dynamic hierarchical reinforcement learning approach offers a robust solution for cost-effective and stable microgrid operation.
• The method demonstrates significant improvements in managing grid-connected microgrids compared to traditional approaches.
• The Java Agent Development Framework (JADE) simulations confirm the practical applicability and validity of the proposed technique.