A prior knowledge-enhanced Transformer model for data anomaly identification and processing in industrial park wastewater treatment plants.

Accurate identification of anomalies and reconstruction of missing data are crucial for the intelligent operation of industrial park wastewater treatment plants (WWTPs). However, traditional statistical and model-based approaches often fall short in addressing anomalies characterized by complex spatio-temporal coupling within WWTP data. Accordingly, this study introduces a novel prior knowledge-enhanced Transformer (PK-T) model. The PK-T model leverages domain-specific knowledge from wastewater treatment within a deep learning architecture, which demonstrates superior performance in the high-precision identification and reconstruction of complex anomalies. The model was evaluated using one year of operational data from two industrial park WWTPs in China. Comparative analysis with conventional methods, including statistical, distance-based, and model-based approaches, reveals that PK-T significantly outperforms these approaches in both anomaly identification accuracy and data reconstruction precision. Remarkably, the model achieves reconstruction accuracies exceeding 90 % for critical water quality parameters. This research provides an innovative solution for robust data management and intelligent operation in industrial wastewater treatment, with considerable practical implications for enhancing the operational efficiency of WWTPs.