Hybrid modelling of nitrogen removal by biofiltration using high-frequent operational data

Abstract

ABSTRACT In this research, a parallel hybrid model is presented for the simulation of nitrogen removal by submerged biofiltration of a very large-size wastewater treatment plant. This hybrid model combines a mechanistic and a machine learning model to produce accurate predictions of water quality variables. The models are calibrated and validated using detailed and quality-controlled operational data collected over a period of 3.5 months in 2020. The mechanistic model is a modified activated sludge model that describes the biological, physical and chemical processes taking place in a biofilm reactor based on the domain knowledge of these processes. A three-layer feed-forward artificial neural network with a rectified linear activation function that aims to reduce the mechanistic model's residual error and then correct its output. The results show how the hybrid model outperforms and significantly reduces the size of the mechanistic model's prediction errors of the effluent nitrate concentration from a relative mean error of 12% (mechanistic model) to 2% (hybrid model) during training. The error on nitrate simulations increases to 8% during hybrid model testing, still significantly lower than the error of the mechanistic model. These results support future operational applications of hybrid biofilm models, such as in digital twins.