Hybrid dynamic modeling of an industrial reactor network with first-principles and data-driven approaches

Abstract

Dynamic modeling for complex, hazardous, difficult-to-operate chemical processes can often be challenging. This paper introduces a dynamic model that utilizes a hybrid approach, combining first-principles and data-driven methodologies, for modeling a complex reactor network comprising seven reactors in a counterflow connection. Specifically, a first-principles model is developed through mechanism analysis for each reactor. Real-time measurements are utilized by an unscented Kalman filter (UKF) to facilitate the co-estimation of both model states and parameters. A quadratic programming optimization is performed to address the physical constraints in the model parameters. Finally, a Seq2Seq neural network is employed in a serial configuration to compensate for the unmodeled dynamics by the first-principles model. The performance of the proposed dynamic model is compared with several other methods on an industrial nitration process under load-changing scenarios. The results demonstrate that our proposed model exhibits superior performance in terms of prediction accuracy and interpretability over other methods.