Abstract: Analyzing the varying operational conditions in accordance to the characteristic parameters of influent water quality in a wastewater treatment plant and based on the multi-model concept, an effluent water quality softsensing model was developed, where the submodel was described by Hammerstein model. With the error BP-like stable learning algorithm and the recursive least square method introduced to learn the multilayer perceptor as nonlinear gain and the ARX model as linear part of Hammerstein model, respectively, the clustering center was adjusted online according to the adjacency between the center and sample. Then, the softsensing method of effluent COD was proposed according to soft switch. The experimental results showed that the clustering centers adjusted online can reflect the varying operational conditions, and that the multi-model softsensing method can offer high accuracy in comparison with operational data.