A new approach to optimizing aeration using XGB-Bi-LSTM via the online monitoring of oxygen transfer efficiency and oxygen uptake rate

Abstract

In wastewater treatment plants (WWTPs), aeration is vital for microbial oxygen needs. To achieve carbon neutrality, optimizing aeration for energy and emissions reduction is imperative. Machine learning (ML) is used in wastewater treatment to reveal complex rules in large data sets has become a trend. In this vein, the present paper proposes an aeration optimization approach based on the extreme gradient boosting–bidirectional long short-term memory (XGB-Bi-LSTM) model via the online monitoring of oxygen transfer efficiency (OTE) and oxygen uptake rate (OUR), thus allowing WWTPs to conserve energy and reduce indirect carbon emissions. The approach uses gain algorithm of XGB to calculate the importance of features and identify important parameters, and then uses Bi-LSTM to predict the target with important parameters as features. Operational data from a WWTP in Suzhou, China, is employed to train and test the approach, the performance of which is compared with ML models suitable for regression prediction tasks (XGB, random forest, light gradient boosting machine, gradient boosting and LSTM). Experimental results show the approach requires only a small number of input parameters to achieve good performance and outperforms other machine-learning models. When OTE and dissolved oxygen (DO) are used as features to predict the alpha factor (αF; since diffusers were used, multiply by the pollution factor F), the R-squared (R2) is 0.9977, the root mean square error (RMSE) is 0.0043, the mean absolute percentage error (MAPE) is 0.0069 and the median absolute error (MedAE) is 0.0032. When the predicted αF and the OUR are used as features to predict the air flow rate of an aeration unit, the R2 is 0.9901, the RMSE is 3.6150, the MAPE is 0.0209 and the MedAE is 1.5472. Using our optimized aeration approach, the energy consumption can be reduced by 23%.