Machine learning-based prediction of biological oxygen demand and unit electricity consumption in different-scale wastewater treatment plants

Abstract

The effluent biochemical oxygen demand (BOD5) and unit energy consumption (UEC) of wastewater treatment plants (WWTPs) are important indicators to measure their performance stability and economy. This study established multiple linear regression (MLR) and machine-learning (ML) prediction models (i.e., back propagation neural network (BPNN), support vector regression (SVR), deep neural network (DNN), and eXtreme gradient boosting (XGBoost)) for effluent-BOD5 and UEC of different-scale WWTPs (i.e., WWTP-A (120,000 m³/d), WWTP-B (53,000 m³/d), WWTP-C (6 000 m³/d)). The initial modeling results show that the comprehensive performance of the ML method is better than the MLR method. To further improve the model prediction performance, the genetic algorithm (GA) is used to optimize the hyperparameters of the BPNN, SVR, and DNN models, and the grid search is used to optimize the hyperparameters of the XGBoost model. The model prediction performance demonstrated that selecting an appropriate prediction method is important for better prediction. In the effluent-BOD5 prediction, the best-performing models for different datasets (i.e., WWTP-A, WWTP-B, WWTP-C and WWTP-MIX) are XGBoost (R2 =0.972, RMSE=0.179), GA-BPNN (R2 =0.853, RMSE=0.223), GA-SVR (R2 =0.933, RMSE=0.244) and GA-SVR (R2 =0.958, RMSE=0.255) respectively. In UEC prediction, the best performing models are XGBoost (R2 =0.709, RMSE=0.013), GA-SVR (R2 =0.712, RMSE=0.035), GA-DNN (R2 =0.858, RMSE=0.077) and GA-DNN (R2 =0.954, RMSE=0.051) respectively. The conclusion of this study also indicated that hybrid modeling of data from different WWTPs is an effective way to improve prediction accuracy. This study provides a relatively novel solution for effluent-BOD5 and UEC prediction and serves as an efficient tool for wastewater treatment and monitoring in WWTPs.