Distributed data-driven modeling of methanol to olefin regenerator based on the artificial neural network

Developing high-quality data-driven models for industrial unit equipment is beneficial for the process control and optimization. To further enhance the model prediction performance, a distributed data-driven modeling approach was proposed in this work with a methanol-to-olefin regenerator as the case study. The approach included data cleaning, operating mode identification, feature selection and sub-models training and package. An MIC-based feature selection method was first proposed to determine feature sets. 40 sets of quasi-raw data were randomly drawn for the model performance evaluation. The results showed that RMSEs and MREs of 7 outputs from the developed model were all smaller than those from the classical data-driven model. The statistically significant test results showed that the prediction performance was improved significantly for 5 outputs, and moderately for 2 outputs. The proposed approach can be applied to other multi-input and multi-output apparatuses, providing a new modeling strategy to promote the smart factory construction.