A novel soft sensor framework based on adaptive LSTM ensemble for robust water quality prediction in wastewater treatment plants

Abstract

Water quality prediction is crucial for efficient operation in wastewater treatment plants (WWTPs). This paper presents a novel soft sensor framework based on an adaptive long short-term memory (LSTM) ensemble for influent water quality prediction. The framework integrates a temporal pattern attention LSTM (TPA-LSTM) model with preprocessing techniques, including the Hodrick-Prescott Filter (HP) for trend extraction and density-based spatial clustering of applications with noise (DBSCAN) for anomaly detection. Missing data are addressed using Newton's Interpolation Method. This combined HP_DBSCAN approach extracts informative features and enhances data quality. The TPA-LSTM model is trained on the preprocessed water quality dataset, enabling accurate prediction of key parameters (e.g., BOD, COD, NH3_N). Comparative analysis demonstrates the superior performance of this hybrid-algorithm model over a single LSTM, highlighting the benefits of the adaptive ensemble approach. The hybrid-LSTM model, in comparison to the single LSTM, has improved the average fitting correlation for seven water quality indicators by 52.7% and reduced the prediction error by 62.1%. Experimental results validate the effectiveness and applicability of the proposed soft sensor for influent quality prediction in WWTPs. This framework offers insights for advancing the use of LSTM networks within wastewater infrastructure, ultimately contributing to improved operational efficiency and reduced costs.