Predicting and interpreting nitrogen removal performance and functional microbial abundance of single-stage partial nitrification and anammox system using machine learning methods.

Machine learning (ML) was employed to simultaneously predict nitrogen removal rate (NRR) and functional microbial abundance of single-stage partial nitrification and anammox (PNA) system. Shapley additive explanations (SHAP) and causal inference were used to analyze the impact of key factors and their optimal ranges. Artificial neural network (ANN) and extreme gradient boosting (XGBoost) have strong predictive abilities for NRR (R2 = 0.94) and functional microbial abundance (R2 ≥ 0.57), respectively. pH and free ammonia (FA) are important factors affecting NRR. To inhibit nitrite oxidizing bacteria (NOB), it was recommended that FA be maintained above 5 mg/L, while O2 be kept below 0.4 mg/L. Candidatus Brocadia-dominated sludge is recommended under low nitrogen (NH4+-Ninf < 200 mg/L) or O2 fluctuation environments, while Candidatus Kuenenia-dominated sludge is recommended under high nitrogen (NH4+-Ninf > 400 mg/L), low temperature (20-30°C), or pH fluctuations (7.4-8.4). These models provide prospects and references for the application of PNA technology.