Double deep reinforcement learning twin-delayed agents for performance improvement of a grid-connected wave energy...

Area of Science:

• Renewable Energy Systems
• Control Systems Engineering
• Artificial Intelligence in Engineering

Background:

• Grid-connected Archimedes Wave Swing (AWS) systems traditionally use proportional-integral (PI) controllers.
• Optimizing energy extraction and minimizing generator losses are key control objectives.
• Existing PI controllers face challenges in dynamic response and efficiency.

Purpose of the Study:

• To introduce deep learning agents trained with the TD3 algorithm as replacements for PI controllers in AWS systems.
• To evaluate a novel hybrid approach combining TD3 agents and PI controllers.
• To benchmark the performance of TD3-based controllers against conventional PI controllers.

Main Methods:

• Two deep learning agents, trained using the TD3 algorithm, were developed.
• The TD3 agents were applied to control generator dq currents and regulate inverter-side voltages.
• A hybrid approach combined two PI controllers with a TD3 agent for the inverter side.
• System performance was analyzed under steady-state and transient fault conditions in MATLAB Simulink.

Main Results:

• The TD3 agents demonstrated effective control over generator dq currents for energy maximization and loss minimization.
• The hybrid TD3-PI controller on the grid side exhibited improved dynamic and steady-state responses.
• Performance benchmarks showed the hybrid TD3-PI agent outperforming the full PI classical configuration.

Conclusions:

• Deep learning agents, particularly the hybrid TD3-PI approach, offer a promising advancement for controlling wave energy conversion systems.
• The proposed TD3-based controllers provide enhanced performance over traditional PI controllers.
• This study validates the reliability and effectiveness of TD3 agents in demanding operational scenarios.