Automated Machine Learning-Based Models for Predicting Effluent Total Nitrogen Concentration of Reclaimed Water in Constructed Wetlands and Precise Regulation of Manganese Ion Dosing Methods

Reclaimed water reuse is a vital strategy for addressing water scarcity, yet elevated total nitrogen (TN) concentrations in reclaimed water remain a major obstacle to its broader implementation. In this experiment, manganese ions (Mn²⁺) at concentrations of 0–8 mg/L were used to enhance the removal efficiency of ammonia nitrogen (NH₄–N), nitrite nitrogen (NO₂–N), nitrate nitrogen (NO₃–N), TN, total phosphorus (TP), and COD in constructed wetlands (CWs). The results showed that Mn²⁺ only improved the removal rates of NO₂–N, NO₃–N, and TN, with the TN removal rate increasing from 11 to 43%. Three different automated machine learning frameworks (Flaml, H₂O AutoML, and AutoGluon) were then applied to predict the effluent TN concentration, with the Flaml model demonstrating the best performance. Under a data set split ratio of 0.8 and a training time of 90 s, the Flaml model achieved an R² of 0.9833, with MAE and RMSE values of 0.145 and 0.182, respectively. Furthermore, the 3D partial dependence plot generated by the optimal model indicated that, while maintaining the effluent Mn²⁺ concentration below 0.1 mg/L, when the influent TN concentration reached its maximum value of 14.84 mg/L, the optimal Mn²⁺ dosing concentration was 6.3 mg/L, resulting in an effluent TN concentration of 4.9 mg/L. This study provides a novel modeling approach for understanding the complex biochemical processes in constructed wetlands for reclaimed water treatment, revealing the dependence between influent and effluent manganese ion concentrations and TN concentrations, and offering a new pathway for the application of artificial intelligence in the field of constructed wetlands.