A Hybrid Learning Method for System Identification and Optimal Control.

Area of Science:

• Engineering
• Control Systems
• Machine Learning

Background:

• System identification and control of nonlinear systems are challenging, especially with limited, low-variability historical data.
• Existing methods often require active system excitation or struggle with closed-loop data, leading to overfitting.
• Data-driven approaches are increasingly important for complex system optimization.

Purpose of the Study:

• To develop a robust, data-driven method for system identification and optimal control of nonlinear systems.
• To address limitations of existing methods when only historical closed-loop data is available.
• To demonstrate the effectiveness of the proposed method on a benchmark system and a real-world building facility.

Main Methods:

• A three-step process involving simulation model creation, neural network-based system physics learning with stopping rules, and reinforcement learning for control.
• Utilizing historical closed-loop data without requiring active system excitation.
• Employing domain randomization and distributed learning techniques within reinforcement learning.

Main Results:

• Successfully identified system models and developed optimal controllers for a pendulum and a large building facility.
• The developed controllers demonstrated superior performance in terms of comfort and energy efficiency compared to benchmark rule-based controllers.
• The method effectively avoids overfitting issues common with closed-loop system identification.

Conclusions:

• The proposed data-driven method offers a viable solution for system identification and optimal control of nonlinear systems using historical data.
• This approach is particularly effective for systems with limited excitation and low-variability control commands.
• The method has practical implications for optimizing energy consumption and comfort in large-scale facilities.