Locally Weighted Principal Component Analysis-Based Multimode Modeling for Complex Distributed Parameter Systems.

Area of Science:

• Engineering
• Computational Science
• Applied Mathematics

Background:

• Global Principal Component Analysis (PCA) is established for Distributed Parameter Systems (DPSs) modeling.
• Parameter variations and multiple operating domains limit the feasibility of global PCA for complex DPSs.

Purpose of the Study:

• Develop a novel multimode spatiotemporal modeling method for large-scale, highly nonlinear DPSs with parameter variations.
• Address the limitations of global PCA by introducing a locally weighted approach.

Main Methods:

• Utilize a finite Gaussian mixture model (FGMM) to decompose spatiotemporal snapshots into Gaussian components.
• Employ Bayesian inference to determine local weights for LW-PCA based on component probabilities.
• Generate local spatial basis functions (SBFs) using PCA on locally weighted snapshot matrices.
• Estimate local temporal models using the Extreme Learning Machine (ELM).

Main Results:

• The developed method effectively models complex DPSs with parameter variations.
• LW-PCA approximates the system using multiple local reduced SBFs, unlike global PCA's single set of SBFs.
• Numerical simulations confirm the effectiveness of the proposed multimode spatiotemporal model.

Conclusions:

• The novel LW-PCA based multimode spatiotemporal modeling method provides a robust solution for complex DPSs.
• This approach enhances modeling accuracy and feasibility in the presence of parameter variations and multiple operating domains.
• The method offers a significant advancement over traditional global PCA techniques for DPS modeling.