Design and Application of a Variable Selection Method for Multilayer Perceptron Neural Network With LASSO.

Area of Science:

• Chemical Engineering
• Computer Science
• Data Science

Background:

• Nonlinear industrial processes often require complex models.
• Existing neural network methods may not efficiently select essential variables.
• Accurate modeling is crucial for process optimization and control.

Purpose of the Study:

• To develop a novel variable selection method for neural networks.
• To improve the modeling of nonlinear industrial processes.
• To achieve higher prediction accuracy and model compression.

Main Methods:

• An iterative two-step approach combining multilayer perceptron and least absolute shrinkage and selection operator (LASSO).
• LASSO with cross-validation determines the shrinkage parameter for variable selection.
• Variables with zero input weights are eliminated, and the process iterates until model accuracy plateaus.

Main Results:

• The proposed method successfully constructs compressed neural network models.
• Demonstrated higher prediction accuracy compared to existing methods in simulations and a crude distillation unit application.
• Effective identification of essential input variables for nonlinear process modeling.

Conclusions:

• The novel variable selection technique enhances neural network performance for industrial applications.
• The method offers a robust way to achieve parsimonious and accurate process models.
• This approach is valuable for optimizing and controlling complex nonlinear systems.