A Novel Dynamic Operation Optimization Method Based on Multiobjective Deep Reinforcement Learning for Steelmaking...

Area of Science:

• Metallurgical Engineering
• Artificial Intelligence
• Process Control

Background:

• Dynamic operation optimization in steelmaking is challenging due to high temperatures and complex reactions.
• Existing methods struggle with real-time adjustments for smelting processes.
• Achieving desired molten steel quality requires precise control of operation parameters.

Purpose of the Study:

• To develop a robust framework for dynamic operation optimization in steelmaking.
• To address the complexities of high-temperature smelting processes.
• To improve the quality control of molten steel production.

Main Methods:

• Application of a deep deterministic policy gradient framework.
• Development of an energy-informed restricted Boltzmann machine for actor-critic networks in reinforcement learning (RL).
• Optimization of neural network (NN) hyperparameters using a multiobjective evolutionary algorithm with a knee solution strategy.

Main Results:

• Experimental validation on real steelmaking data demonstrates the method's practicability.
• The proposed approach shows significant advantages and effectiveness over existing methods.
• The system successfully meets specified molten steel quality requirements.

Conclusions:

• The developed deep reinforcement learning framework offers an effective solution for dynamic steelmaking optimization.
• The energy-informed restricted Boltzmann machine enhances decision-making in complex industrial processes.
• This research contributes to advancing intelligent manufacturing in the steel industry.