Multi-strategy modeling integrating kinetics mechanism of cracking and pyrolysis and unsupervised dual-stage attention long and short-term memory network

The fluid catalytic cracking process utilizing the dual-riser reactors (MIP-LTAG) holds significant importance in the development of petrochemical enterprises. It aims to reduce fuel consumption while increasing output. Consequently, modeling for the production process is an essential task. However, traditional methods struggle to accurately describe the complex reaction mechanisms involved in the cracking/pyrolysis dual reaction pathways. Additionally, due to the coupling of variables and insufficiency of dynamic characteristics, capturing multi-variable spatio-temporal dependencies remains challenging. This paper focuses on key indicators such as product yield and carbon emissions within the core reaction-regeneration unit of the target technological process. A lumped kinetic mechanism model is constructed to balance the reaction pathway. Variational mode decomposition (VMD) is employed to perform decomposition of the coupled variables. The unsupervised dual-stage attentional long short term memory model (UDA-LSTM) is utilized to capture multi-scale characteristics. To leverage these advantages, this paper designs three hybrid model for collaborative optimization of multi-objective predictions. Finally, the effectiveness of the proposed hybrid modeling framework is validated through an actual industrial production case. The predicted mean squared error (MSE) of the main product yield does not exceed 0.2, and the constructed process model supports real-time monitoring of the production process by refineries.