Dual-stage soft sensor-based fault reconstruction and effluent prediction toward a sustainable wastewater treatment plant using attention fusion deep learning model

Abstract

Soft sensor-based monitoring of wastewater treatment plants (WWTPs) is crucial for ensuring stable operation, maintaining strict environmental standards, and minimizing economic losses. However, faulty measurements of independent variables generate inaccurate data, which affects the reliability of the developed soft sensor. Therefore, this study proposes a two-stage modeling approach to reconstruct faulty measurements and predict effluent water quality parameters using attention-fusion techniques with convolutional deep learning model. In the first stage, faulty measurements are reconstructed using an attention-fusion autoencoder (AFAE) model. In the second stage, the reconstructed data are then fused into an attention convolutional neural network (ACNN) to provide real-time predictions of effluent parameter concentrations. In the reconstruction stage, the AFAE model achieved a superior fault reconstruction performance for a malfunctioning dissolved oxygen sensor with R² value of 0.9909. In the subsequent stage, the ACNN model exhibited superior predictive capabilities for effluent parameter concentrations, reducing residual error by 57.2 % compared to the faulty data scenario. Consequently, the aeration energy saving was improved by 18.4 % with the sustainable environmental discharge of the effluent. The proposed two-stage AFAE–ACNN model-based soft sensor can simultaneously calibrate malfunctioning sensors and accurately predict effluent concentrations, providing smart operational strategies for sustainable WWTPs.