Operation Optimization Decision-Making of Aluminum Electrolysis Process Using Offline Reinforcement Learning.

Area of Science:

• Industrial Process Optimization
• Artificial Intelligence
• Chemical Engineering

Background:

• Aluminum electrolysis generates complex datasets with diverse and risky policies.
• Offline reinforcement learning faces challenges in ensuring safety and feasibility for industrial operations.
• Existing methods struggle with distribution shift in multicategory behavioral policies.

Purpose of the Study:

• To develop an offline multiobjective reinforcement learning method with multicategory policy constraint for aluminum electrolysis operation optimization.
• To enable learning of optimization policies that exceed behavior policies while adhering to safety requirements.
• To address the distribution shift problem in multicategory behavioral policies.

Main Methods:

• Proposed an offline multiobjective reinforcement learning framework with multicategory policy constraint.
• Utilized a mixture Gaussian variational autoencoder (GMVAE) for behavior cloning and policy constraint.
• Designed an actor-critic architecture with separate critics for operational performance and safety.
• Introduced a safety critic network to enforce industrial safety requirements.

Main Results:

• The proposed method learned an optimization policy superior to behavior policies.
• The learned policy successfully met stringent industrial safety requirements.
• Experimental results on real-world aluminum electrolysis data showed superior performance compared to other offline reinforcement learning algorithms.
• Effectively alleviated the distribution shift problem on multicategory behavioral policies.

Conclusions:

• The developed offline multiobjective reinforcement learning approach is effective for safe operation optimization in aluminum electrolysis.
• The method provides a robust solution for complex industrial scenarios with safety constraints.
• The approach demonstrates significant advantages over existing offline reinforcement learning algorithms in this domain.