Model construction and application for effluent prediction in wastewater treatment plant: Data processing method optimization and process parameters integration.

Abstract Wastewater treatment based on the activated sludge process is complex process, which is easily affected by influent quality, aeration time and other factors, leading to unstable effluent. Facing increasingly stringent sewage discharge standards in China, it is necessary to build a prediction model for early warning of effluent quality. In this study, nine machine learning algorithms were adopted to construct models for the prediction of effluent Chemical Oxygen Demand (COD). In order to improve the prediction accuracy of the models, model optimization was conducted by introducing the hysteresis condition [Hydraulic Retention Time (HRT) of 18 h], data processing method (K-FOLD) and process parameters [dissolved oxygen (DO), sludge return ratio (SRR) and mixed liquid suspended solids (MLSS)]. Results showed that both K-Nearest Neighbour (KNN) and Gradient Boosting Decision Tree (GBDT) displayed excellent prediction effects, the best results of MAPE, RMSE and R2 were 7.34%/1.29/0.92(COD, KNN). The optimized models were further applied to the prediction of effluent total phosphorus (TP), total nitrogen (TN) and pH. The MAPE/RMSE/R2 were 7.43%/0.92/0.93(TN, GBDT), 17.81%/0.19/0.99(TP, KNN), 0.53%/0.16/0.99 (pH, KNN) respectively, indicating high prediction accuracy. The change and comparison of modeling conditions provide a new insight to wastewater prediction models. In addition, this study is close to the actual application scenario of WWTPs operation and management, also laying a foundation for the reverse regulation of energy saving and consumption reduction of wastewater treatment plants (WWTPs).