Modelling approach for high flow rate in wastewater treatment operation

The vast majority of models developed for wastewater treatment and receiving water systems have been of the distributed-parameter and state-space (lumped-parameter) forms. On the other hand, most control system design methods and, for that matter, methods of system identification refer to the black box class of models and in particular the time series models. In the current study two modelling techniques of the black box class of models were used to model the data collected from full-scale treatment operations at the Gold Bar Wastewater Treatment Plant (GBWWTP), the largest plant in the Edmonton area (Alberta, Canada). An artificial neural network was trained to make short-term predictions of the quantity of wastewater entering the plant during storm events using rainfall data collected from eight gauges covering the parts of the city that are serviced by combined sewers. After training, the model was able to generalize very well when tested against an unseen set of data. Transfer function time series models were used to model the quality data collected from a primary sedimentation tank at the plant. The models were able to make hourly predictions of the total suspended solids and chemical oxygen demand in the primary effluent. The presented models have the predictiveness and adaptiveness requirements needed for models that could be utilized as part of a real-time control system. Key words: dynamic modelling, artificial neural networks, transfer-function models, wastewater inflow, primary sedimentation.