A hybrid approach to soft sensor development for distillation-in-series plant under input data low variability

Abstract

This paper presents a hybrid approach for integrating fundamental process knowledge with measurement data to soft sensor (SS) development with improved estimation capability. Measurement data from sensors are collected and used as inputs for a first-principles model to emulate the data close to restrictions of the operating regulations, thus addressing a low variability problem of the inputs. Next, variables from measurement data and results of the first-principles modeling are combined to extend the training dataset for SSs, which become of a hybrid type in nature. To improve an estimation capability, a cascade-forward neural network and algorithm for alternating conditional expectation for nonparametric SS development was used. It was shown that the estimation capabilities of the developed SS can be improved by extending the training dataset with first-principles model data approximating the upper and lower limits of the process regime, the size of which in total does not exceed 21% of industrial data alone. As a result, the designed hybrid SS demonstrates a better efficacy in predicting quality index of the targeted distillation product with significantly reduced mean absolute error.