Learning-based MPC of sampled-data systems with partially unknown dynamics.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Dynamical Systems

Background:

• Real-world control systems often feature time-varying parameters and irregular data sampling, complicating accurate modeling and stability guarantees.
• Existing model predictive control (MPC) methods struggle with these uncertainties, limiting their applicability.
• Accurate system identification and robust control are critical for reliable operation.

Purpose of the Study:

• To propose a novel learning-based model predictive control (LMPC) method for sampled-data control systems with partially unknown dynamics.
• To address challenges posed by time-varying parameters and irregular data sampling.
• To ensure system stability and boundedness in uncertain environments.

Main Methods:

• A neural ordinary differential equation (NODE) is employed to learn unknown time-varying dynamics from irregularly sampled data.
• The learned dynamics model is integrated into a sampled-data MPC framework.
• Gronwall-Bellman inequality is utilized to derive conditions for guaranteeing ultimate boundedness.

Main Results:

• The proposed LMPC method effectively learns and adapts to unknown system dynamics.
• Quantitative stability analysis confirms the system's ultimate boundedness.
• The method's applicability is demonstrated through two practical examples.

Conclusions:

• The developed LMPC offers a robust solution for controlling systems with partially unknown and time-varying dynamics.
• The integration of NODE and MPC provides a powerful framework for handling irregular sampling.
• The approach enhances the reliability and stability of sampled-data control systems in real-world applications.