Mechanism- and data-driven based dynamic hybrid modeling for multi-condition processes

Abstract

In process industries, the complexity and variability of working conditions make it challenging to accurately measure product quality. While data-driven models have developed rapidly, they often overlook the underlying physical or chemical mechanisms. To address this, we propose a hybrid modeling approach that combines mechanism- and data-driven methods. Historical and current working condition data are processed through a hidden layer to extract features. The partial differential equation is discretized and approximated using the forward Euler method to derive mechanism-based quality variable values. These values are then combined with real data through a weighted mix to create a new label for dynamic regression. Additionally, a domain adaptation regularization term is introduced to align the distributions of different working conditions. Through analyses of three process industry datasets, we demonstrate that this method can predict unmeasurable variables with reasonable accuracy and exhibits stronger generalization ability compared to pure data-driven models.