Domain adaptation network with uncertainty modeling and its application to the online energy consumption prediction of ethylene distillation processes

Real-time monitoring of quality index especially energy consumption is of great significance to improve energy efficiency and save energy. In ethylene distillations, the energy consumption cannot be measured online due to the lack of corresponding online analytical instruments and soft sensing is an effective technology to solve the problem. Additionally, ethylene industrial processes often operate under multiple conditions due to the change of the working conditions, in which the insufficient data will lead to poor prediction performances of the soft sensing model. Therefore, in this paper, a novel soft sensing method based on transfer learning framework is proposed to deal with the above problems, which consists of a feature extractor and a predictor. Concretely, to reduce the difference of domain distribution in the feature extractor, the lower bound estimation of Wasserstein distance is proposed as a metric of the difference. In the predictor, to mitigate the performance degradation caused by the domain difference, the adaptive features for predictor are constructed by the source domain features and uncertainty modeling of the quantified domain difference, which is measured through the proposed manifold distance. Therefore, an adaptive transfer learning framework is proposed to reduce effect of the domain difference, named as Domain Adaptation Network with Uncertainty Modeling (DANUM). Finally, a soft sensing method based on DANUM is proposed to predict quality index and energy consumption in ethylene distillation processes, which shows the effectiveness of our method.

Automated summary

This paper proposes a soft sensing method based on transfer learning framework to address the issues of real-time monitoring and prediction of energy consumption in ethylene distillation processes. By reducing the domain distribution difference in the feature extractor, constructing adaptive features for the predictor, and quantifying the uncertainty of domain difference modeling, the effectiveness and accuracy of the soft sensing model are significantly improved.