Modeling nutrient removal by membrane bioreactor at a sewage treatment plant using machine learning models

Abstract

This study developed machine learning (ML) models to predict nutrient removal using an anaerobic-anoxic-oxic membrane bioreactor (A2O-MBR). An extreme gradient boosting (XGBoost) model was applied using a grid search strategy (Grid-XGBoost) to predict the removal of nutrients, including ammonium (NH4), total phosphorus (TP), and total nitrogen (TN). The models were validated against a commonly used multilayer perceptron (MLP) neural network. The input parameters were divided into operating conditions, including dissolved oxygen, oxidation-reduction potential, and mixed liquor suspended solids. These conditions were also partitioned based on influent characteristics such as NH4, TN, TP, total organic content, chemical oxygen demand, and suspended solids. A total of nine models were developed for each ML technique using the operating conditions and influent characteristics as separate datasets and combining them for each target nutrient. It was observed that using only operating conditions or influent characteristics as input parameters for XGBoost and MLP yielded poor results. Moreover, a significant improvement in the predictive efficacy of the model was observed when all parameters for the target nutrient removal predictions were considered. The prediction of NH4 by the XGBoost model had the highest R2 values of 0.763, 0.814, and 0.876 when the operating conditions, influent characteristics, and combined dataset were used as input parameters, respectively. Overall, the ensemble XGBoost model demonstrated better performance than the MLP model in all cases. However, the performance of both the models was found to be inadequate for predicting TN and TP removal in any scenario. The proposed XGBoost model is a reliable and robust ML technique for predicting NH4 removal, which may contribute to decision-making in advance to improve the efficacy of an A2O-MBR system.