Hierarchical deep reinforcement learning for self-adaptive economic dispatch

Area of Science:

• Power Systems Engineering
• Artificial Intelligence
• Renewable Energy Integration

Background:

• Modeling power system uncertainty with large-scale intermittent renewables (wind, photovoltaic) is challenging due to volatility.
• Deep reinforcement learning (DRL) offers adaptive strategies but faces sparse rewards and high-dimensional issues in such systems.

Purpose of the Study:

• To develop an efficient method for power system economic dispatch under uncertainty.
• To address the limitations of standard DRL in large-scale, uncertain power systems.

Main Methods:

• A hierarchical deep reinforcement learning (HDRL) scheme was designed.
• The HDRL approach decomposes the problem into a global stage (RL agent) and a local stage (heuristic algorithm).

Main Results:

• The proposed HDRL scheme demonstrated efficiency in solving power system economic dispatch problems.
• HDRL successfully adapted to system uncertainties and volatility, improving online decision-making speed.

Conclusions:

• HDRL is an effective strategy for optimizing power system operations with significant renewable energy integration.
• The method enhances adaptability and decision-making speed in complex, uncertain power system environments.