RSSM-Based Virtual Sensing and Sensorless Closed-Loop Control for a Multi-Temperature-Zone Continuous Crystallizer.

Area of Science:

• Chemical Engineering
• Control Systems
• Artificial Intelligence

Background:

• Precise temperature control is vital for product quality and energy efficiency in continuous crystallizers.
• Industrial crystallizers present challenges due to nonlinear dynamics, strong coupling, and unobservable states from sensor limitations.
• Traditional control strategies struggle with partial observability and model uncertainty.

Purpose of the Study:

• To develop a Model-Based Reinforcement Learning (MBRL) framework for precise temperature control in multi-zone continuous crystallizers.
• To address challenges posed by unobservable states and model uncertainty using offline historical data.
• To create a high-fidelity digital twin that functions as a real-time virtual sensor.

Main Methods:

• A Recursive State Space Model (RSSM) was developed as a digital twin and virtual sensor to infer unobservable system states.
• A multi-objective reward function was designed to balance tracking error, stability, and control cost.
• The MBRL framework utilized solely offline historical data for training.

Main Results:

• The virtual sensor demonstrated exceptional long-term stability and high fidelity in predicting unobservable states.
• The trained MBRL agent significantly outperformed traditional Proportional-Integral-Derivative (PID) and Model Predictive Control (MPC) controllers.
• Over 67% improvement in temperature tracking accuracy and over 93% reduction in control action costs were achieved.

Conclusions:

• The proposed MBRL framework with a virtual sensor effectively addresses partial observability in continuous crystallizers.
• The virtual sensing capability enables precise state estimation for robust control policy optimization.
• The approach enhances system operation, leading to improved energy efficiency and product quality.