Collaborative twin actors framework using deep deterministic policy gradient for flexible batch processes.

Area of Science:

• Process Control
• Artificial Intelligence
• Chemical Engineering

Background:

• Batch processing is efficient but challenging to control with flexible conditions.
• Traditional batch-to-batch learning control struggles with limited prior information.
• Optimizing performance in dynamic batch systems requires advanced control strategies.

Purpose of the Study:

• To develop a novel deep reinforcement learning (DRL) approach for flexible batch process control.
• To address limitations of traditional methods in handling varying operating conditions and initial states.
• To enhance control policy generation and ensure safe operation in complex batch systems.

Main Methods:

• Proposed a collaborative twin-actor-based deep deterministic policy gradient (CTA-DDPG) method.
• Utilized sequential actor-critic networks with a shared critic for offline meta-policy exploration and online performance enhancement.
• Incorporated policy integration and spatial-temporal experience replay for robust transfer and efficient learning.

Main Results:

• CTA-DDPG demonstrated effective control policy generation for flexible batch processes.
• The method ensured safe operation across varying trial lengths and initial conditions.
• Evaluations on numerical examples and an injection molding process confirmed superior performance.

Conclusions:

• The CTA-DDPG method offers a superior solution for flexible batch process control.
• This DRL approach effectively overcomes limitations of traditional learning control strategies.
• The proposed method achieves desired control outcomes in complex, dynamic industrial settings.