Modeling of a mainstream partial nitrification/anammox process through a hybrid theoretical-machine learning approach

Abstract

Model simulations are vital in optimizing and predicting the performance of biological wastewater treatment, especially for processes involving slow-growing bacteria. However, data records often include missing, invalid, or infrequent measurements of parameters, compromising prediction accuracy. This study used a hybrid theoretical-machine learning approach to address these issues. By leveraging the stoichiometry and kinetics, missing values were calculated in limited data sets, which were then analyzed through machine learning algorithms to reveal hidden microbial interactions. The model was validated with data from a pilot-scale partial nitritation/anammox fluidized bed membrane bioreactor (PN/A FMBR) with saline sewage. The model demonstrated strong prediction performance, with random forest outperforming other algorithms with correlation coefficients of 0.89, 0.72, and 0.80 for ammonium, nitrite, and nitrate data sets, respectively, when compared to actual values. Training sets containing 73 to 88 same-day values reached acceptable predicting performance. The results also revealed that microbial synergy in nitrogen transformation, particularly in the partial denitrification from nitrate to nitrite linked to Anammox in responding to a low DO supply, was evident in this PN/A FMBR. Additionally, key parameters, including temperature, pH, and specific microbiomes, were identified as critical for predicting PN/AFMBR performance, highlighting significant microbial interactions that warrant further investigation.