Input variable selection for model-based production control and optimisation

Abstract

For model-based production control and optimisation, it is crucial to properly identify those input variables that have the strongest influence on production performance. This way, production operators can focus only on the relevant variables, and production control problems can be reduced. In order to identify previously unknown relationships among the production variables, hidden knowledge in historical production data needs to be explored. In the article, two decisive steps are considered. First, an input variable selection methodology, typically applied for selecting model regressors, is applied. Next, the appropriateness of the selected inputs and their manipulative strength is validated by an operating-space-based controllability analysis. To use the most appropriate input variable selection approach, different input selection methodologies are compared with synthetic data sets. Moreover, a case study of Tennessee Eastman process is applied to demonstrate a complete input variable selection procedure for model-based production control and optimisation.