Temperature-aware and feature-augmented modeling framework for mechanism-informed effluent quality prediction in cold-region wastewater treatment plants.

In cold regions, accurate prediction of effluent quality in wastewater treatment plants (WWTPs) remains challenging due to temperature-induced process heterogeneity and nonlinear interactions among influent indicators, environmental factors, and effluent quality. This study proposes a Temperature-Aware and Feature-Augmented Modeling (TAFAM) framework to improve the prediction of COD, TN, and TP. This framework incorporates temperature-driven stage division to capture stage-specific process behavior and augments input features using knowledge-derived interaction features reflecting load and ratio relationships. A dataset comprising 14,808 samples collected from a cold-region WWTP was used for model development. The effectiveness of TAFAM was evaluated for nine machine learning algorithms under consistent experimental conditions. Results showed that TAFAM significantly enhanced the capture of nonlinear relationships, with the ensemble methods demonstrating superior performance. Compared with baseline models, TAFAM reduced RMSE by 6.98%-17.89%, and MAE by 12.17%-23.39%, while increasing R2 by 7.90%-22.02%. The SHAP and PDP analysis suggested stage-specific mechanistic heterogeneity, identifying potential threshold effects for key parameters (e.g., NH₃-N/TN ratio 0.84 for COD, pH window 6.9-8.0 for TP). The TAFAM framework not only improved model adaptability and interpretability but also bridged the gap between data-driven modeling and mechanistic understanding, facilitating a paradigm shift from experience-based to data-driven, mechanism-informed decision-making. This offered a novel methodology for process understanding and operational guidance in cold-region WWTPs, with potential applicability through adaptive adjustments.