Modeling Nutrient Dynamics in Sequencing Batch Reactor

Abstract

The use of artificial neural networks (ANN) for modeling complex processes is an attractive approach that has been successfully applied in various fields. However, in many cases the use of an ANN alone may be inadequate and inaccurate when data are insufficient, because the ANN black-box model relies completely on the data. As a result, a hybrid model consisting of a simplified process model (SPM) and a neural network (residual model) is used in the present study for developing a dynamic model of sequencing batch reactor systems. The implemented SPM model consists of only five discrete rate equations and an ANN is added to the SPM in a parallel connection. Both the SPM and the ANN receive influent chemical oxygen demand (COD), total kjeldahl nitrogen (TKN), PO43- and NH4+ data and timer output signals (for phase control) as inputs. The SPM output provides a preliminary prediction of the dynamic behavior of the PO43- and NOx- concentrations. The outputs of the trained ANN compensate for the output errors of the SPM model. The hybrid model output of the final predictions of the process states is obtained by summing the outputs from both the SPM and ANN. Successful application of such a hybrid model is demonstrated.