This paper presents a systematic methodology based on the application of artificial neural networks for sequencing batch reactor (SBR) identification. The SBR is a fill-and-draw biological wastewater technology, which is specially suited for nutrient removal. The proposed approach makes optimal use of the available data during the training stage and it is aimed at achieving high generalization ability. For this purpose, a wide range of experimental conditions, including different solids retention times and influent characteristics, has been used. The methodology is successfully applied to develop a soft-sensor for monitoring a laboratory-scale SBR operated for enhanced biological phosphorus removal. The main interest is the utilization of the soft-sensor to determine the optimal length of the SBR stages within each cycle according to the actual process requirements. Note that SBRs are normally operated with constant pre-defined duration of the stages, thus, resulting in low efficient operation. Data obtained from the on-line electronic sensors installed in the SBR and from the control quality laboratory analysis have been used to develop the optimal architecture of two different ANNs. The ANNs were trained for on-line prediction of phosphorus (P) concentration in the SBR. One ANN uses only inexpensive and reliable on-line measurements as input data and the other one also includes as input the previous P measurement (lag-1), thus considering the quality variable dynamics. The latter ANN can be used to overcome the delay introduced by the measurement procedure of phosphorus concentration. The results have shown that the developed models can be used as efficient and cost-effective predictive tools for the system analysed, since they accurately reproduced the phosphorus behaviour in the SBR. (C) 2008 Elsevier Ltd. All rights reserved.