Accurate dynamic models are critical in chemical process engineering for control, optimization, and safety, yet many processes exhibit nonlinear dynamics that linear models fail to capture. This thesis investigates how modern machine-learning-based system identification methods compare with classical approaches on nonlinear process benchmarks and examines which models offer the best trade-off between accuracy, complexity, and robustness. Two benchmark systems are studied: a simulated two-tank single-input single-output (SISO) process and a quadruple-tank multiple-input multiple-output (MIMO) system, considered in both minimum-phase and nonminimum-phase operating regimes. The input-output datasets were generated by simulating the established white-box models for both the two-tank and quadruple-tank systems. All candidate models are formulated in a Nonlinear AutoRegressive with eXogenous input (NARX) framework and trained in series–parallel mode, then evaluated in free-run simulation. The study compares linear Polynomial ARX and nonlinear Polynomial NARX models, a feedforward Neural NARX model, and a CatBoost-based NARX ensemble. The results show that models using only linear regressors are fundamentally limited, whereas nonlinear models achieve substantially higher fidelity. Polynomial NARX and CatBoost NARX with degree = 2 regressors provide clearly better accuracy on the two benchmarks, with Polynomial NARX achieving the highest accuracy overall. The Neural NARX model proves highly sensitive to configuration, performing reasonably with linear regressors but becoming unstable with nonlinear inputs. Overall, the thesis demonstrates that structured nonlinear models such as Polynomial NARX and tree-based ensemble methods like CatBoost can provide high-fidelity dynamic models suitable for demanding applications. In the context of process safety, these findings support the use of accurate data-driven models as a basis for advanced monitoring, anomaly detection, and model-based control in chemical process systems.
Modelli dinamici accurati sono fondamentali nell’ingegneria dei processi chimici per il controllo, l’ottimizzazione e la sicurezza; tuttavia, molti processi presentano dinamiche non lineari che i modelli lineari non riescono a catturare. Questa tesi indaga come i moderni metodi di identificazione di sistemi basati sul machine learning si confrontino con gli approcci classici su benchmark non lineari e valuta quali modelli offrano il miglior compromesso tra accuratezza, complessità e robustezza. Sono studiati due sistemi di riferimento: un processo a due serbatoi a singolo ingresso e singola uscita (SISO) simulato e un sistema a quattro serbatoi a ingressi e uscite multipli (MIMO), considerati sia in regime a fase minima sia non a fase minima. I dataset ingresso–uscita sono stati generati simulando i modelli white-box consolidati per entrambi i sistemi. Tutti i modelli candidati sono formulati nel framework NARX (Nonlinear AutoRegressive with eXogenous input) e addestrati in configurazione serie–parallelo, per poi essere valutati in simulazione in free-run. Lo studio confronta modelli ARX polinomiali lineari e NARX polinomiali non lineari, un modello Neural NARX feedforward e un ensemble NARX basato su CatBoost. I risultati mostrano che i modelli che utilizzano solo regressori lineari sono intrinsecamente limitati, mentre i modelli non lineari raggiungono una fedeltà sostanzialmente superiore. I modelli Polynomial NARX e CatBoost NARX con regressori di grado = 2 forniscono un’accuratezza nettamente migliore sui due benchmark, con il Polynomial NARX che ottiene l’accuratezza complessiva più elevata. Il modello Neural NARX risulta altamente sensibile alla configurazione: si comporta in modo soddisfacente con regressori lineari, ma diventa instabile con regressori non lineari. Nel complesso, la tesi dimostra che modelli non lineari strutturati, come il Polynomial NARX, e metodi ad ensemble basati su alberi, come CatBoost, possono fornire modelli dinamici ad alta fedeltà adatti ad applicazioni impegnative. Nel contesto della sicurezza di processo, tali risultati supportano l’uso di modelli data-driven accurati come base per il monitoraggio avanzato, il rilevamento di anomalie e il controllo basato su modello nei sistemi di processo chimico.
System identification using machine learning methods
ABDOLAHZADE, KASRA
2024/2025
Abstract
Accurate dynamic models are critical in chemical process engineering for control, optimization, and safety, yet many processes exhibit nonlinear dynamics that linear models fail to capture. This thesis investigates how modern machine-learning-based system identification methods compare with classical approaches on nonlinear process benchmarks and examines which models offer the best trade-off between accuracy, complexity, and robustness. Two benchmark systems are studied: a simulated two-tank single-input single-output (SISO) process and a quadruple-tank multiple-input multiple-output (MIMO) system, considered in both minimum-phase and nonminimum-phase operating regimes. The input-output datasets were generated by simulating the established white-box models for both the two-tank and quadruple-tank systems. All candidate models are formulated in a Nonlinear AutoRegressive with eXogenous input (NARX) framework and trained in series–parallel mode, then evaluated in free-run simulation. The study compares linear Polynomial ARX and nonlinear Polynomial NARX models, a feedforward Neural NARX model, and a CatBoost-based NARX ensemble. The results show that models using only linear regressors are fundamentally limited, whereas nonlinear models achieve substantially higher fidelity. Polynomial NARX and CatBoost NARX with degree = 2 regressors provide clearly better accuracy on the two benchmarks, with Polynomial NARX achieving the highest accuracy overall. The Neural NARX model proves highly sensitive to configuration, performing reasonably with linear regressors but becoming unstable with nonlinear inputs. Overall, the thesis demonstrates that structured nonlinear models such as Polynomial NARX and tree-based ensemble methods like CatBoost can provide high-fidelity dynamic models suitable for demanding applications. In the context of process safety, these findings support the use of accurate data-driven models as a basis for advanced monitoring, anomaly detection, and model-based control in chemical process systems.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Kasra_Thesis_01.pdf
accessibile in internet per tutti
Dimensione
5.65 MB
Formato
Adobe PDF
|
5.65 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Kasra_Executive Summary_02.pdf
accessibile in internet per tutti
Dimensione
3.17 MB
Formato
Adobe PDF
|
3.17 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/247496