Ensemble Non-Gaussian Local Regression for Industrial Silicon Content Prediction

Abstract

Due to the complicated characteristics of modeling data in industrial blast furnaces (e.g., nonlinearity, non-Gaussian, and uneven distribution), the development of accurate data-driven models for the silicon content prediction is not easy. Instead of using a fixed model, an ensemble non-Gaussian local regression (ENLR) method is developed using a simple just-in-time-learning way. The independent component analysis is utilized to handle the non-Gaussian information in the selected similar data. Then, a local probabilistic prediction model is built using the Gaussian process regression. Moreover, without cumbersome efforts for model selection, the probabilistic information is adopted as an efficient criterion for the final prediction. Consequently, more accurate prediction performance of ENLR can be obtained. The advantages of the proposed method is validated on the online silicon content prediction, compared with other just-in-time-learning models.