Anna Mikola

Graphical abstract showing that calibrated CFD of a full-scale 12.6 m aeration basin was converted into a deep-learning surrogate, reducing dissolved-oxygen prediction time from 1,856 s to 61–67 ms with 0.11–0.40 mg/L error. This study aimed to develop a fast surrogate model for spatial dissolved oxygen prediction in a deep, full-scale activated sludge aeration basin and to evaluate its suitability for operational use. A two-phase water-air computational fluid dynamics model was built in COMSOL Multiphysics 6.3 for a basin with a water depth of 12.6 m using a realizable k-epsilon turbulence model. A computational fluid dynamics-informed deep neural network was then trained on computational fluid dynamics-generated datasets using Bayesian hyperparameter optimization. Model performance was evaluated against measured dissolved oxygen data from two independent 24-hour field campaigns. Bootstrap analysis yielded 95% confidence intervals of 0.11–0.40 mg/L for the MAE and 6.0–17.7% for the normalized root mean square error. The surrogate reduced evaluation time from 1,856 seconds to 61–67 milliseconds, corresponding to an average speed-up of about 27,900-fold. These results indicate that the framework can provide real-time spatial dissolved oxygen estimates and show strong potential for plant-specific monitoring and aeration control support.

Digital twin models offer great potential for process improvements in wastewater treatment plants (WWTPs). Such models require a constant real-time input data feed from the physical process. Collecting these data is challenging, especially in the harsh conditions in the headworks of the process. In this study, data-driven models and process and sewer system expertise were combined to design soft-sensors for primary effluent COD and NH4-N prediction. Ordinary least squares regression and the seasonal autoregressive integrated moving average model with exogenous variables were tested using flow rate and suspended solids concentration as model input. An excellent NH4-N prediction was achieved, and the prediction accuracy was further improved by implementing process-insight-driven weights. The tested models were able to achieve either good COD estimation accuracy or effectively capture the variability in the target data. However, achieving both simultaneously remained challenging, with or without weights. Simulation tests using the calibrated process model demonstrated that the developed soft-sensors were able to provide real-time predictions leading to goodness-of-fit in simulations comparable to or better than that achieved using laboratory data influent quality.