Neural Network-Based Hammerstein Model Identification of a Lab-Scale Batch Reactor.

Area of Science:

• Chemical Engineering
• Artificial Intelligence
• Control Systems

Background:

• Batch reactor processes, such as acrylamide polymerization, exhibit complex nonlinear dynamics.
• Accurate system identification is crucial for effective controller design.
• Traditional Hammerstein model identification can be computationally intensive.

Purpose of the Study:

• To develop and compare two neural network-based methods for Hammerstein model identification.
• To identify nonlinear systems efficiently for simplified controller design.
• To explore advanced machine learning applications in process control.

Main Methods:

• Gradient-based backpropagation algorithm for training a multilayer neural network.
• Extreme learning machine (ELM) for training a single hidden-layer feedforward network representing the nonlinear block.
• Direct parameterization of Hammerstein model blocks using neural network weights.

Main Results:

• Both neural network approaches successfully identified the Hammerstein model for the batch reactor process.
• The ELM-based method demonstrated efficient training without gradient calculations.
• The identified models facilitate easier linear and nonlinear controller design.

Conclusions:

• Neural network-based Hammerstein model identification provides an effective approach for complex nonlinear systems.
• The ELM method offers a computationally efficient alternative for parameter estimation.
• Future work includes implementing machine learning-based nonlinear model predictive control.