Implications of the Training Data Set on the Performance of Property Estimators

摘要

In this paper the implications of the characteristics of the training data set on the estimator performance are investigated. A simulated crude distillation unit is chosen as case study, and different training sets are used to build linear steady-state product property estimators. A comparison of estimators based on different secondary variables and trained on different data sets is presented. It is shown that only if specific inputs to the inferential model are adopted (consistent inputs) the estimator performances are almost independent of the training data set. If inappropriate inputs are used, the estimator performance deteriorates and may change significantly with the training set used.