Automatic control of simulated moving bed process with deep Q-network.

Area of Science:

• Chemical Engineering
• Artificial Intelligence
• Process Control

Background:

• Simulated moving bed (SMB) process control is complex due to nonlinear partial differential-algebraic equations and discrete events.
• Active product purity constraints are susceptible to violations from disturbances.

Purpose of the Study:

• To apply a data-based deep Q-network (a model-free reinforcement learning method) for optimal control of SMB processes.
• To train a near-optimal control policy for complex dynamic systems using artificial intelligence.

Main Methods:

• Utilized a deep Q-network for off-line training of control policies using simulated data.
• Implemented a large penalty for constraint violations and logic-based switching control to manage extract and raffinate purities.
• Leveraged parallel numerical simulations for efficient data generation.

Main Results:

• The trained deep Q-network achieved fast on-line computation of control inputs within SMB process time limits.
• Indirect imposition of state constraints via penalties and switching control proved effective.
• Demonstrated the advantages of deep reinforcement learning for SMB process control.

Conclusions:

• Deep reinforcement learning offers a viable solution for the optimal control of challenging SMB processes.
• The combination of deep Q-learning with logic-based control enhances constraint satisfaction and exploration efficiency.
• AI-driven control strategies show significant promise for complex chemical engineering applications.