Comparative Study of Machine Learning and System Identification for Process Systems Engineering Dynamics.

Area of Science:

• Process Systems Engineering (PSE)
• Data-Driven Modeling
• Machine Learning (ML)

Background:

• Traditional system identification methods face challenges in complex dynamical systems.
• Modern machine learning (ML) offers potential for improved data-driven modeling in PSE.
• Integrating ML into PSE requires robust frameworks and evaluation strategies.

Purpose of the Study:

• To benchmark traditional system identification versus ML models for PSE applications.
• To evaluate the efficacy of MLOps-inspired tools for system identification tasks.
• To provide insights into optimal model selection and performance evaluation in PSE.

Main Methods:

• Utilized AutoSID, an automated framework inspired by Machine Learning Operations (MLOps).
• Compared 12 diverse model architectures (system identification, ML, deep learning) across 11 PSE case studies.
• Employed 4 model search/hyperparameter optimization algorithms and 3 model selection criteria.

Main Results:

• Model selection is critical for effective system identification.
• Bayesian optimization with Tree-structured Parzen Estimators (TPE) is effective for balanced model selection.
• K-fold cross-validation is a robust performance evaluation metric; information criteria are efficient for large datasets.
• ML models with balanced complexity, such as tree ensembles, demonstrate superior predictive accuracy and computational efficiency.

Conclusions:

• MLOps-inspired workflows can enhance system identification in PSE.
• Balanced complexity ML models are recommended for superior performance in PSE.
• Findings offer actionable insights for PSE practitioners in model selection and evaluation.