Five-day biochemical oxygen demand (BOD5) is one of the key parameters which is widely used to evaluate the biological and chemical evaluation of effluents from wastewater treatment plants. This paper proposes a novel method that can predict online using BOD5 by synthesizing an online version of Gustafson-Kessel (GK) algorithm and least squares support vector machines (LS-SVM). The clustering algorithm can reduce required number of clusters, form more complex shape clusters, and get better modeling performance. Moreover, an online sparse LSSVM with time window is proposed to reduce the computation time and storage space. The GK-LSSVM of the method of the soft-sensor model was proposed to predict BOD5 concentration. The results indicate that the proposed method can not only improve prediction accuracy but also efficiently decrease model's update frequency, comparing to the cases using with that of different methods.