Causal-informed domain adaptation with cross-attention for predicting anomalous effluent quality fluctuations in...

Accurate prediction of anomalous effluent quality fluctuations in wastewater treatment plants (WWTPs) is essential for safeguarding aquatic environments, and the main challenge lies in limited time-series data and imbalanced samples. However, existing transfer learning methods have difficulty identifying anomalous fluctuations in this heterogeneous system due to influent variability and process diversity. To address this, a causal-informed temporal domain adaptation model enhanced with cross-attention (CA-TDA) is proposed for time-series prediction. Unlike conventional methods that focus on distribution alignment, CA-TDA dynamically captures and aligns causal structures across heterogeneous domains. Temporal causal features are extracted via a variational autoencoder (VAE), upon which cross-attention identifies key driving factors for effective cross-domain adaptation. Experiments show that even when the target domain contains only 8.35 % of source samples, CA-TDA achieves superior prediction of multiple effluent quality indicators (EQIs), with an average R² of 0.9669 and a peak R² of 0.9820 for TN. Compared to baselines, F1- and F2-scores for consecutive anomalous fluctuations are improved by 29-87 % and 30-89 %, respectively. These results confirm that CA-TDA effectively captures complex, non-stationary fluctuations, substantially enhancing the accuracy and robustness of cross-domain effluent anomaly prediction, demonstrating broad application potential.