Using LSSVM model to predict the silicon content in hot metal based on KPCA feature extraction

Abstract

To overcome the difficulty that silicon content in hot metal can not be effectively controlled in ironmaking process due to lack of real-time on-line instrumentation, a prediction method is proposed by combining the Kernel Principal Component Analysis(KPCA) with the Least Square Support Vector Machine(LSSVM). Using KPCA as a preprocessor of LSSVM to extract the principal features of original data and employ the 10-fold cross validation to optimize the parameters of LSSVM. Then LSSVM is applied to proceed silicon content regression modeling. KPCA can denoise the input data and capture the high-ordered nonlinear principal components in input data space, and with LSSVM we can establish a prediction model between the featured principal components and the primary variable for the silicon content in iron making processes. The data of the model are collected from No.6 Blast Furnace in Baotou Iron and Steel Group Co. of China. The results show that the LSSVM model based on KPCA feature selection has higher accuracy and better tracking performance compared with LSSVM or PCA-LSSVM models, so the proposed method can satisfy the requirements of on-line measurements of silicon content in hot metal.