Prediction of heavy diesel T95 using just in time learning models

Abstract

One of the middle distillates of atmospheric distillation column is Heavy Diesel (HAD). T95 is the temperature, at which 95% volume of a sample is boiled, and it is the main controlled variable so accurate T95 predictions are required for a satisfactory performance of the model predictive control (MPC) algorithm, in which online T95 predictions are used to determine control actions. In the current thesis, just in time learning (JITL) methodology is used on historical process data to develop soft sensors for real time predictions of HAD T95. Local models are constructed using samples located in the neighborhood of a query point, and the constructed model is used for prediction of the response variable. Using 47 process variables in the historical data, three main groups of predictive models are constructed for HAD T95. In the first group of models, various subsets of variables, which are assumed to carry the highest information on variation of T95, are included into static and dynamic models. While there is no time lag between the selected input variables and the predicted quality variable in static models, previous day’s T95 values (response variable) are included in dynamic models, as in autoregressive exogenous (ARX) input modeling. In the second group of models, least-squares (LS), partial LS (PLS), and subset regression via stepwise regression methods are employed on a predictor set, which consists of seven “most important” process variables. JITL models are evaluated with respect to various reference data selection methods, reference set size, window size and neighborhood size. The best model of this group is found to have predictive root means square error (RMSE) and mean absolute error (MAE) statistics equal to 5.66 and 4.23 oC, respectively. In the last group of models, interaction and quadratic predictor terms are included in the JITL model, and neighboring samples are selected a different subset of predictors. Using this method, RMSE and MAE of prediction statistics are decreased to 4.77 and 3.82 oC, respectively. This, to our knowledge, is the first time predictor and neighbor selection predictor subsets are separated from each other in the literature, and this seems a promising method in constructing soft-sensors for industrial applications.