$k$-Step Look-Ahead Active Concurrent Learning-Based Dual Control of Exploration and Exploitation for...

Area of Science:

• Control Systems Engineering
• Machine Learning
• Renewable Energy Systems

Background:

• Auto-optimization in systems with unknown references and environments presents significant challenges.
• Balancing parameter estimation and optimal reference tracking is crucial for effective system control.
• Existing methods often struggle with convergence speed and reliance on specific excitation conditions.

Purpose of the Study:

• To introduce a $k$-step look-ahead active concurrent learning-based dual control of exploration and exploitation (KSLCL-DCEE) framework.
• To address challenges in auto-optimization by inherently balancing parameter estimation and optimal reference tracking.
• To achieve faster convergence and relax the need for persistent excitation in control systems.

Main Methods:

• Developed a KSLCL-DCEE framework with inner and outer loops utilizing future cost function gradients.
• Implemented a $k$-step look-ahead mechanism to generate control commands based on estimated reference trajectories.
• Introduced active concurrent learning with a modified learning rate for accelerated convergence.

Main Results:

• Demonstrated the capability of KSLCL-DCEE to effectively balance exploration and exploitation.
• Achieved faster convergence rates compared to existing methods by modifying the learning rate.
• Provided a comprehensive stability analysis confirming the robustness of the KSLCL-DCEE framework.

Conclusions:

• The KSLCL-DCEE framework offers a robust solution for auto-optimization in systems with unknown dynamics.
• The proposed method shows significant performance improvements and faster convergence, validated by numerical studies.
• Successful application on photovoltaic (PV) arrays highlights the practical utility of the KSLCL-DCEE algorithm.