Neural network identification in nonlinear model predictive control for frequent and infrequent operating points...

Area of Science:

• Chemical Engineering
• Control Systems
• Artificial Intelligence

Background:

• Nonlinear Model Predictive Control (NMPC) struggles with highly nonlinear dynamic processes due to model inaccuracies at infrequent operating points (IOPs).
• Existing nonlinear models, identified from frequent operating points (FOPs), lead to tracking errors and potential instability when processes deviate.
• Dynamic processes often exhibit short-term shifts from FOPs to IOPs, challenging traditional control strategies.

Purpose of the Study:

• To develop a novel optimal identification algorithm for training NMPC nonlinear models.
• To enhance the accuracy and robustness of NMPC in highly nonlinear dynamic systems, particularly during transitions to IOPs.
• To ensure stable and precise control across both FOPs and IOPs.

Main Methods:

• A novel optimal identification algorithm is proposed, tailored to the plant's nonlinearity.
• A Multi-Layer Perceptron (MLP) neural network is employed as the nonlinear model.
• The MLP is trained to accurately represent system behavior at FOPs while maintaining acceptable performance at IOPs.

Main Results:

• The proposed algorithm successfully trains an MLP to describe nonlinear dynamic system behavior accurately.
• The trained model demonstrates acceptable performance even in infrequent operating points.
• Validation on a nonlinear dynamic pH neutralization process shows significant improvements in simulation and implementation.

Conclusions:

• The developed optimal identification algorithm effectively addresses NMPC challenges in highly nonlinear dynamic processes.
• The method improves model accuracy and control system stability across varying operating conditions.
• Simulation and implementation results confirm the algorithm's practical effectiveness and robustness.