A practical method for maintaining the validity of hybrid AI process models without retraining
Journal
Computers and Chemical Engineering
Journal Volume
204
Start Page
109432
ISSN
00981354
Date Issued
2026-01
Author(s)
Liu, Hsiao-Te
Fang, Ming-Chun
Lee, Hao-Yeh
Hsieh, Cheng-Ting
Hua, Tzu-Chieh
Lin, Shih-Chieh
Lee, Chih-Lung
Huang, Tzu-Hsien
Chou, Wei-Ti
Abstract
A challenge with AI models is that model validity usually deteriorates over time as properties of the process gradually change. For example, catalyst activity in reactors may decrease due to poisoning or thermal degradation, the heat transfer coefficient in heat exchangers may decrease due to fouling, and the tray efficiency in distillation columns may decrease due to plugging. Usually retraining is required to restore the model prediction performance, which is costly and time-consuming. To address this problem, a practical hybrid AI framework is proposed that incorporates equipment parameters such as tray efficiency as input variables. This allows the model accuracy to be restored by adjusting these parameters, which significantly reduces the need for the time-consuming and expensive process of retraining the model. The industrial applicability of this method is demonstrated using an industrial process for coke oven gas (COG) scrubbing and light oil recovery. The results show that the error of using the AI model developed in 2024 to predict the data for 2021 is up to 37.55%. By adjusting the equipment parameters, the error can be reduced to 3.77%. This method can effectively solve the problem of model degradation over time and can be applied to most chemical processes.
Subjects
Artificial intelligence
Cyber-physical systems
Neural networks
Physics-guided machine learning
Smart manufacturing
SDGs
Publisher
Elsevier Ltd
Type
journal article