Performance prediction of the full-scale bardenpho process using a genetic adapted time-delay neural network (GATDNN)

Wastewater treatment systems are characterized by large temporal variability of inflow, variable concentrations of components in the incoming wastewater to the plant, and highly variable biological reactions within the process. The behavior of observed process variables within a wastewater treatment plant (WWTP at a certain time instant is the combined effect of various processes initiated at different moments in the past. This is called a time-delay effect in the system. Due to the nature of strong nonlinear mapping, neural networks provide advantages as a modeling and identification tool over a structure-based model. However, the determination of the architecture of the artificial neural networks (ANNs) and the selection of key input variables with a time delay is not easy. in our research, a genetic adapted time-delay neural network (GATDNN), which is a combination of time-delay neural network(TDNN) and genetic algorithms(GAs), was developed and applied to the full-scale Bardenpho advanced sewage treatment process. In a GATDNN, a three-step modelling procedure was performed: (1) selection of significant input variables to maximise the predictive accuracy for each specific output; (2) finding a suitable network topology for the ANN-based process estimator; (3) sensitivity analysis. The results demonstrate that the modelling technique presented using a GATDNN provides a valuable tool for predicting the outputs with high levels of accuracy and identifying key operating variables. This work will permit the development of a reliable control strategy thus reducing the burden of the process engineer.