Because a single model cannot represent the characteristics of the complex industrial process in varying operating ranges, we propose an online modeling scheme for multiple models 0 This scheme inc1udes the recognition mechanism of operating range, the local models and the combination mechanism for multiple models. The recognition mechanism analyzes the operating range according to its characteristic variables and adjusts the center of operating range according to similarity degrees. The loca1 model is actua11y a Hammerstein model which is the seria1 connection of a wavelet neural network with a stable learning algorithm and an ARX model. The combination mechanism ca1culates the weighted sum of the outputs of local models, and online adjusts the centers of operating range to reflect the variation characteristics of the operating range. A stable learning algorithm of parameters improves the prediction accuracy and the adaptation ability. This method is implemented in a wastewater treatment process to measure the concentration of the chemical oxygen demand (COD). Experimental results show that 由is modeling scheme can obtain satisfactory effect in varying operating ranges.