Long Short-Term Memory Neural Networks for Modeling Dynamical Processes and Predictive Control: A Hybrid...

Area of Science:

• Chemical Engineering
• Artificial Intelligence
• Control Systems

Background:

• Accurate modeling of complex chemical processes like polymerization is crucial for efficient control.
• Traditional models often struggle with the nonlinear dynamics inherent in polymerization reactions.
• Data-driven approaches like neural networks offer potential but may lack physical interpretability.

Purpose of the Study:

• To introduce a novel physics-informed hybrid neural network (PIHNN) model.
• To develop a computationally efficient model predictive control (MPC) algorithm utilizing the PIHNN.
• To validate the effectiveness of the PIHNN modeling and MPC control strategies.

Main Methods:

• Developed a PIHNN integrating first-principle physics with Long Short-Term Memory (LSTM) neural networks.
• Employed a fuzzy logic-based data fusion block to combine physics-based and data-driven components.
• Designed a computationally efficient MPC algorithm leveraging the capabilities of the PIHNN model.

Main Results:

• The PIHNN model demonstrated highly accurate simulation results for the polymerization reactor.
• The MPC controller, based on the PIHNN, achieved excellent control quality.
• The hybrid approach successfully merged physical insights with data-driven learning.

Conclusions:

• The proposed PIHNN offers a robust and accurate modeling solution for polymerization processes.
• The developed MPC strategy provides effective and efficient control for the simulated reactor.
• This hybrid approach represents a significant advancement in applying AI to chemical process control.