Dynamic study of inferential sensors (Neural Nets) in quality prediction of crude oil distillation tower side streams

摘要

Prediction of properties by statistical methods, and especially by neural networks, is a field that is developing extremely fast. For refinery units, there exist many examples of application that use real plant data (usually, hourly averages) to calibrate neural models. The models are able to predict interesting properties like distillation end points for the products of a crude oil distillation tower. Nevertheless, few examples are known where a dynamic study is performed to highlight the importance of variable evolution along time. The objective of this work is to show how first-principles based dynamic simulators can be used to take into account the dynamics of a crude oil distillation column. In few words, dynamic data is generated with a commercial simulator (Hysys.Plant~+), this data is used to generate the neural net model and, later, the goodness of the proposed methodology is checked against dynamic plant test runs. This proposed mixed procedure combines the use of real plant data to generate the steady state gain and the dynamic simulation data to develop the finite impulse response models. The corrected dynamic models should offer better predictive characteristics than models with variables not conveniently delayed.