Online optimal modeling of LS-SVM based on time window

Least squares support vector machines (LS-SVM) is a perfect model-learning algorithm with good accuracy and high speed. Previously, many researches have been done on this algorithm in static modeling problems, but not in dynamic ones. In this paper, we try to solve these problems. Through researches, we propose a new kind of online modeling algorithm based on time window in LS-SVM and use it for modeling of complex nonlinear processes. The purpose of this paper is to show its powerful identification performances. The paper first presents the main mechanism of LS-SVM, and then discusses the optimization algorithm of time window and describes the key action of Karush-Kuhn-Tucker (KKT) optimization condition to this algorithm. The current feature of the model has strong relationship with L updated data. KKT optimization condition decides whether to do the retraining at each updating procedure and avoids unnecessary recalculations. LS-SVM provides great help for increasing the speed during the online modeling. Finally, this algorithm is applied to solving a multivariable modeling problem of a typical complex process in calcination kiln. The simulation results show the good prospect of this algorithm on dynamic identifications of complex nonlinear processes.