Toward smart wastewater treatment plants: a novel data-driven sludge blanket model based on stochastic differential equations

Abstract

ABSTRACT A novel data-driven stochastic state space system for modeling and forecasting the sludge blanket height in secondary clarifiers is presented. The model is trained on sensor measurements of the sludge blanket height and uses as inputs (1) the clarifier sludge mass inflow rate, and (2) the clarifier recycle flow rate. The model's prediction accuracy is evaluated on data from two Danish wastewater treatment plants, for a summer and a winter month, by means of root-mean-square errors and compared with a persistence model. The model consistently outperforms the persistence model in the summer, but only one plant performs well in the winter month. The worst performing plant is challenging to model due to data quality issues and problematic (uneven and time-varying) flow distributions to the clarifiers. This led us to conclude that the best performance and stability is seen to require high data quality and well-controlled flow distribution. In summary the model achieves, in almost all cases, prediction error reductions in the order of 30–50% and 0.1–0.4 m in relative and absolute terms when compared with the predictions from a persistence model.