Key performance index estimation based on ensemble locally weighted partial least squares and its application on industrial nonlinear processes

Recent decades have witnessed a trend that soft sensing, instead of hard sensing, has been extensively applied to estimate the key performance indices under the circumstances that practical measurements are hardly to be achieved at a reasonable cost. However, due to the existence of nonlinearities and time-varying characteristics in the practical industrial processes, the conventional soft sensor models probably suffer from severe performance degradations when the original designed models are mismatched. Although many novel methodologies have been employed to alleviate this problem, each of them merely focuses on certain aspect of model features, a comprehensive framework combining these features is needed. Therefore, this study proposes an online predictive methodology based on an integration of ensemble learning based on a novel adaptive locally weighted partial least squares. Specifically, sub-models established on the respective dataset are generated by moving window model, time difference model and just-in-time learning model for the sake of different properties in processes. The effectiveness of the proposed model is validated on the practical nonlinear processes represented by a benchmark simulation model No.1 (BSM1), in wastewater treatment plants (WWTP), and a real industrial catalytic reforming process.