The probabilistic discriminative time-series model with latent variables and its application to industrial chemical process modeling

Abstract

Probabilistic models with latent variables have caught extensive attention in industrial modeling since they can predict the distribution of variables of interest. However, the traditional probabilistic model is a generative model in which some assumptions are made to calculate the logarithmic likelihood and evaluate the optimal probabilistic model with the maximum logarithmic likelihood. Inaccurate assumptions may degrade the prediction performance of the variables of interest of the generative model. Thus, we propose the probabilistic discriminative model with fewer assumptions to improve the discriminating performance of the model. The industrial hydrocracking process is a complex chemical process in which the proper prior knowledge of the distribution is hard to obtain. To avoid introducing unnecessary and inaccurate assumptions, a probabilistic discriminative model based on the latent variable time-series model is constructed. The proposed method is then verified for the industrial hydrocracking process with lower root mean square error by 16.9%.