Toward explainable and generalizable data-driven modeling in real wastewater treatment plants: Utilizing...

Data-driven models have increasingly been used as useful tools for process simulation in urban wastewater treatment plants (WWTPs), but their lack of interpretability and limited generalization hinder their application in practical engineering scenarios. In this study, a hybrid neural network incorporating an attention-based encoder and a long short-term memory module was proposed for data-driven modeling of the anaerobic-anoxic-oxic (A2O) process in real WWTPs. Compared with several widely used baseline deep learning architectures for time-series prediction, the proposed method demonstrated improved accuracy and efficiency in predicting dissolved oxygen (DO) concentrations in the oxic zone. The model was further analyzed using a bidimensional interpretable deep learning framework, enabling global interpretability analysis across temporal and feature levels. This bidimensional perspective provided new insights for optimizing model architecture and feature selection. Three practical scenarios characterized by stepped water temperature distributions were defined, and local interpretability methods were employed to enhance understanding of the individual and interactive effects of aeration rate and water temperature on DO within each scenario. The results indicate that the interaction mechanisms among core variables affecting DO vary significantly across different seasonal conditions. Therefore, it is necessary to design scenario-specific aeration strategies to optimize real-time control operations in WWTPs. Additionally, the transfer learning method, incorporating dynamic system similarity, significantly enhanced the model's generalization capability. This suggests that transferring and retraining models across similar A2O systems may be an effective approach for improving model reuse and addressing the "cold start" challenge in newly constructed WWTPs.