A Mechanism-Data Codriven Hybrid Modeling Framework for WWTPs to Achieve Reliable Simulation at System-Level

Building high-precision wastewater treatment plant (WWTP) models is important to ensure effluent safety and control operational costs. However, physics-based models suffer from idealized assumptions and unmeasurable parameters, leading to insufficient simulation accuracy. In contrast, data-driven models have limited practicality due to poor interpretability and weak extrapolation capability. In this study, we developed a mechanistic-data codriven hybrid modeling framework that embeds first-principles knowledge as physical constraints, guiding data-driven components to compensate for critical unmodeled dynamics. The pollutant removal process can be simulated using differential equations representing biochemical reaction mechanisms through the construction of a mechanistic model based on Activated Sludge Model No. 2d (ASM2d), thus confining outputs to defined ranges. Long short-term memory (LSTM) was selected as a suitable data-driven component, showing the ability of capturing hidden patterns within mechanistic errors sequence. The constructed ASM2d–LSTM hybrid model achieved high accuracy in predicting effluent chemical oxygen demand (COD), total phosphorus (TP), and total nitrogen (TN), with test set R2 values of 0.92, 0.91, and 0.94 and mean absolute percentage error (MAPE) of 3.88%, 3.26%, and 8.03%, respectively. It outperformed exclusively data-driven models in dynamic tracking, precision, and generalization. Ablation experiments and SHapley Additive explanation (SHAP) analysis interpreted ASM2d-LSTM’s decision logic. Specifically, the hybrid framework integrates richer online monitoring data, overcoming mechanistic model input constraints and enhancing prediction. Overall, our work establishes hybrid modeling as a guiding paradigm for achieving precise WWTP modeling, thereby providing foundational technical support for achieving low-carbon plant operations.