Knowledge-guided machine learning with an operating time index for selectivity prediction and optimization in industrial acetylene hydrogenation

Catalyst deactivation is a persistent challenge in industrial acetylene hydrogenation, where ethylene selectivity is a key performance indicator. This study introduces a knowledge-guided machine learning framework that incorporates an operating time index (OTI) to capture the catalyst aging effects. The OTI consists of an integer component denoting the cycle length and a decimal component representing the normalized progression within each cycle. By embedding this index into a deep neural network (OTI-DNN), we achieved substantial improvements in plant-scale prediction accuracy compared with conventional Deep neural network (DNN) and Long short-term memory (LSTM) models. Using five years of operational data from a naphtha cracking center, the OTI-DNN reduced the normalized root mean squared error for ethylene selectivity by 24% compared with a conventional DNN without OTI. Integration of the model with Bayesian optimization further enhanced the process performance, yielding a 21.4% increase in ethylene production and a 14.9% improvement in annual profit by optimizing the catalyst regeneration scheduling. This study demonstrates that OTI provides a practical surrogate for catalyst deactivation, enabling accurate data-driven prediction and optimization without relying on complex dynamic models. The proposed framework can be extended to other catalytic processes, offering a generalizable approach for industrial chemical engineering.