Online sensor-guided hybrid ordinary differential equation-based model of nitrification dynamics in a full-scale...

Accurate modeling of nitrification using mechanistic activated sludge models (ASMs) remains challenging because biological dynamics are typically calibrated through laborious and expert-dependent procedures. In this study, ASM No. 3 (ASM3) was integrated with an artificial neural network (ANN) to form a supporting data-enhanced hybrid ordinary differential equation (H-ODE) framework for a 69-day simulation of a full-scale municipal wastewater treatment plant (WWTP) operation. The ANN was provided with an indirect microbial indicator of nitrification activity extracted from online ammonium-nitrogen (NH₄-N) data. This key input served as supporting data to compensate for unmodeled nitrification dynamics, resulting in a root mean square error of 0.24 for NH₄-N prediction during testing, compared with 0.49 obtained using a conventional H-ODE. In addition, physical interpretability (PI), quantified by the PI index representing mechanistic contribution, increased from 0.70 ± 0.01 (H-ODE) to 0.75 ± 0.08. Notably, in the context of the present framework and dataset used, conventional serial parameter calibration provided only marginal improvement in predictive performance after online sensor-guided process compensation, suggesting that the proposed strategy can compensate for dominant unmodeled nitrification dynamics while maintaining high PI. Online sensor-guided process compensation based on the default ASM3 parameters also improved the predictive performance while maintaining high PI. Although the predictive accuracy for oxidized nitrogen remains constrained by the lack of denitrification activity indicators, the proposed framework demonstrates the potential of online ammonium sensor data to support physically interpretable compensation of nitrification dynamics in hybrid modeling in full-scale municipal WWTPs.