Temperature prediction model for rotary kiln based JITL with regularized extreme learning machine

Sintering process of the rotary kiln is often characterized by multivariable coupling, serious nonlinearity and obvious time varying, which makes difficult for accurate prediction of burning zone temperature. Thus, a new soft sensor model called JRELM based JITL (just in time learning) is developed to forecast temperature. First, considering redundancy between input variables, PCA (Principal component analysis) is utilized for feature selection. Then, in the JITL framework, local RELM (Regularized Extreme Learning Machine) model with strong nonlinear capability and high online efficiency is developed for temperature prediction. Furthermore, moving window strategy is adopted to update historical database and thus enhance model adaptive capability. Based on the industrial data from an alumina enterprise in China, the experimental result comparison demonstrates that estimation performance of the JRELM performs better than conventional adaptive soft sensor algorithms, as well as higher online efficiency.