A three stage attention enabled stacked deep CNN-BiLSTM (ASDCBNet) model for end-to-end monitoring of wastewater treatment plant

Abstract

Rapid urbanization and industrialization have drastically increased wastewater generation, leading to a decline in water quality and threatening both ecosystems and public health. With over 40% of the global population lacking access to clean water, the role of wastewater treatment plants (WWTPs) has become crucial in removing contaminants and safeguarding the environment. However, traditional WWTPs face challenges including high operational costs, manual monitoring dependencies, and inefficiencies in real-time quality control. To address these challenges, this work proposes an automated WWTP monitoring system ASDCBNet powered by deep learning. The system comprises three integrated components: a CNN-BiLSTM-based inflow forecasting model, an attention-enabled sensor health monitoring module, and a CNN-based outflow water quality classification model. The proposed model achieved high forecasting accuracy, with RMSE and MAE values of 95.23 m³/day (1.36%) and 80.23 m³/day (1.15%), respectively, on inflow volumes ranging from 7000 to 10,000 m³/day. Furthermore, the model achieved an exceptionally low mean absolute percentage error of just 0.05%, highlighting its ability to effectively handle variability in the data, ensuring high accuracy in inflow volume forecasting. The model outperforms sensor health monitoring and prediction with an average accuracy of 96.6%, and finally, outflow analyses have reported the prediction accuracy as 98.7%. The model has demonstrated excellent overall performance in statistical analysis, with a Bias of 0.02, a high correlation coefficient of 0.98, a Nash–Sutcliffe Efficiency of 0.85, and a low Thiel’s U-statistic of 0.12. The model's practical application in real-world WWTPs can enhance operational efficiency, reduce manual labor, and improve water quality management by providing accurate, real-time insights.