Predicting propylene loss with inferential model development using design of experiments (DOE) and historical data
Inferential models are a highly researched topic as the science of digital automation becomes more prevalent as information is in abundance. Well-developed inferred models can augment the use of analyzers in steady state processing and highly correlated ones can even replace online analytics. The use of design of experiments (DOE) inferred models with historical process data and a rigorous plant simulator can reduce the case study duration while achieving a high degree of accuracy. This paper uses surface response and full factorial models as the first step in model development, and then uses actual historical plant data to create a well]defined inferred property. The main effect interactions need to be analyzed and identified to determine their statistical significance. This step is important because the interaction factors will not be left out of the final model if significant. Thorough analysis shows the identified model to be robust but exhibits orthogonality in the uncoded units equation. The model was reduced to the main effects to remove orthogonality. Once complete a strong model was identified with good success and application. After identification, the model was tweaked to the historical data for better accuracy. This method proves successful in reducing plant step tests and case study durations to develop a good cost effective correlated model.^
Applied Mathematics|Engineering, Chemical
Wheeler, Jeffrey Allen, "Predicting propylene loss with inferential model development using design of experiments (DOE) and historical data" (2014). ETD Collection for University of Texas, El Paso. AAI1564704.