From black-box prediction to transparent process control: A physics-informed interpretable AI framework for total...

Accurate forecasting of effluent total nitrogen (TN) is essential for stable operation and compliance-risk management in wastewater treatment plants (WWTPs). However, nonlinear biochemical reactions, multiscale temporal dependencies, and weak physical coherence in purely data-driven models still limit prediction reliability. This study proposes a physics-informed dual-attention BiLSTM-Kolmogorov-Arnold network framework, termed PI-DABiLSTM-KAN, for accurate and interpretable effluent TN prediction. LightGBM-SHAP was used to screen process-relevant variables from high-dimensional monitoring data, and ASM1/ASM3-informed descriptors were constructed to represent nitrogen-transformation potential. Feature and temporal attention were introduced to capture key variables and long-range temporal dependencies, while a KAN decoder enhanced nonlinear mapping. Soft constraints derived from mass conservation and Monod kinetics were embedded into a two-stage adaptive training strategy to improve physical plausibility. The framework was evaluated using long-term online monitoring data from a full-scale WWTP in China and an independent external WWTP validation dataset. On the main test set, PI-DABiLSTM-KAN achieved R² = 0.9796, RMSE = 0.3580 mg/L, and MAE = 0.0968 mg/L, with less than 0.8% R² degradation under noise perturbation. It maintained R² > 0.95 for 1-60 h forecasts and achieved R² = 0.9610 on the external validation dataset, with stable seasonal performance across spring, summer, autumn, and winter. Physical-consistency analysis showed that the adaptive physics strategy reduced Any violation from 4.21% to 1.87%. SHAP-CCM analysis further distinguished model-sensitive proxy variables from dynamically coupled candidate drivers, identifying Coarse_Screen_Pump_Flow as a controllable variable with dynamic-coupling support. Counterfactual regulation analysis indicated that coordinated adjustment of hydraulic loading, influent pH, and aerobic pH could reduce high-risk TN predictions within the 0-15 mg/L compliance-risk screening interval. These results suggest that PI-DABiLSTM-KAN provides a process-plausible decision-support framework for effluent TN prediction and operational screening in WWTPs.