Calibration-Augmented and Mechanism-Driven Deep Learning Hybrid Framework for Modeling Actual Distillation Processes

Accurate prediction models are pivotal to improving the production efficiency and ensuring product quality in distillation processes. Traditional mechanism-based models neglect real-world fluctuations, while data-driven models suffer from noise and overlook chemical constraints, leading to inaccurate data and diminished performance. Therefore, a hybrid framework that embeds the mechanism-based model and data calibration into deep learning is proposed to leverage the complementary capabilities of both methodologies. The framework solves the problem of insufficient data accuracy of deep learning models by data calibration, including nonparameter regression, liquid level correction, and a robust estimator. It also takes thermodynamic constraints into account by integrating the mechanism-based models with the convolutional neural network (CNN), thereby capturing dynamic relationships between variables and efficiently predicting the key process parameters. The calibration-augmented and mechanism-driven CNN hybrid framework achieves exceptional predictive performance, validating the effectiveness of complex distillation modeling, further offering a novel insight into mechanism-based and data-driven hybrid paradigms for a digital twin in intelligent factories.