(339l) Development of Data Based Prediction Model for Yield and Composition of Distillate from Vdu Process

Vacuum distillation unit(VDU) is equipment for lube base oil(LBO) plant which fractionates the unconverted oil(UCO) varying with the carbon structure and carbon number. The distillates separated by VDU have their different physical properties and it can be applied to produce the various lubricants. The most important properties to determine the product quality are kinematic viscosity, specific gravity, and viscosity index, so we measure those properties in the commercial plant to set the operating condition. However, it is hard to establish an online monitoring system due to the time-consuming measurement procedure. As a breakthrough, this work developed the neural network-based model to predict the yield and 2-dimensional(2D) composition of each distillate using the 1-dimensional(1D) composition of UCO and distillates, and the operating condition of VDU. Herein, 1D composition indicates the concentration distribution with the carbon number regardless of the structure, on the other hand, 2D composition means the concentration distribution with both number and structure.

For the yield prediction neural network model, the input variables are the 1D composition of UCO and the operating condition of VDU. The output variables are yields of each distillate. It is hard to expect the fine performance of the model with simple augmentation of UCO composition and the operating condition due to its heterogeneity of the data. Therefore, the raw input matrix is preprocessed using the 1D composition of distillate to reflect the operating condition indirectly. The input matrices are newly organized through extracting the range of D5% to D95% of each distillate from the raw UCO 1D composition matrix, and the output matrices only include the corresponding yield of distillate respectively. Accordingly, the neural network models as many as the number of distillates were finally constructed. To develop the model for the 2D composition of distillate, we design the model for the 1D composition of distillate preferentially. The input variable, in this case, contains the 1D composition of UCO, and the operating condition, and the output variable contains the 1D composition of each distillate. After the model is designed, we can compute the yields and 1D compositions of each distillate when any UCO is separated under any operating condition. To infer the 2D compositions of distillate with this information, there should be an additional assumption that the carbon structure distributions of distillate are same with the UCO ones. The final model can predict the yield and 2D composition of distillate with high accuracy. This work is quiet different from the conventional deconvolution method because the operating condition affects the result of the separation. So the operating condition is properly reflected to the neural network model. The result of this study can be utilized to develop a virtual LBO plant that predicts the quality of the final product according to the UCO and operating condition by combining with a model that computes the physical properties of distillate.