(145b) Predictive Modeling for Product Quality in Polymerization Process

Achieving consistent and on-spec product quality is critical for chemical production processes. Final product samples are taken at the end of processes to measure key product qualities in the lab. Polymerization processes have substantially long and varying residence time between process adjustments, product sampling and lab analysis, depending on production rates. The long lag time means that there is a high potential for long periods of sub-prime material production before getting back on-spec. Sub-prime material inventory reduces a plant’s effective production capacity and incurs added cost for storage, segregation and blending, which negatively impacts working capital performance and profitability. Hence, there is a need to develop predictive analytic models that provide real-time prediction of final product quality, enable process adjustment without delay, and minimize the production of off-spec products.

This paper presents a Machine Learning-based predictive model developed by Honeywell Connected Plant for a chemical plant located in Texas to provide real-time prediction of final product viscosity. The prediction model was developed using process data, laboratory data, advanced data analytics coupled with deep process understanding. To prevent overfitting, Machine Learning leverages cross-validation, splitting the data set into a training data set and a validation data set to ensure models are both highly accurate and generalize well to new situations. It was found that the selection of relevant contributing variables and proper data aggregation and preparation are critical for developing an accurate and meaningful data-driven predictive model. The effects of model input variables are consistent with polymerization reaction mechanism. The final selected model successfully predicts viscosity for multiple product grades.