A Robust Reinforcement Learning Control Method for Uncertain Process Industry Based on Knowledge-Constrained...

Area of Science:

• Chemical Engineering
• Control Systems Engineering
• Artificial Intelligence

Background:

• Process industries face challenges optimizing continuous manufacturing due to resource and energy constraints.
• Reinforcement learning (RL) shows promise for control but is hindered by uncertain disturbances affecting reliability.
• Existing methods struggle to accurately model and mitigate disturbances in complex industrial processes.

Purpose of the Study:

• To develop a robust reinforcement learning (RRL) control method for optimizing process indicators in uncertain environments.
• To address the impact of inherent and external uncertainties on the reliability of industrial process control.
• To enhance the performance of control strategies in the process industry despite unpredictable disturbances.

Main Methods:

• Proposed a knowledge-constrained adversarial perturbation method for robust RL (RRL).
• Developed a reaction atmosphere indicator surrogate model to quantify inherent uncertainty.
• Introduced a dynamic state perturbation set with an update policy and an external uncertain time series generation method.

Main Results:

• The RRL method effectively characterizes uncertain disturbances by perturbing observed states.
• The surrogate model quantifies inherent uncertainty, and dynamic perturbation ensures rationality.
• Case validation in zinc electrowinning demonstrated enhanced control performance under uncertainty.

Conclusions:

• The proposed knowledge-constrained adversarial perturbation RRL method significantly improves control reliability in uncertain industrial scenarios.
• Composite modeling and surrogate models are effective for handling process uncertainties.
• The approach offers a viable solution for optimizing continuous manufacturing processes with enhanced robustness.