Explainable Temporal Deep Learning for Sustainable Urban Wastewater Infrastructure in Cognitive Cities

Effective wastewater treatment is essential for achieving environmental sustainability, particularly given the growing global population and the increasing volume of wastewater produced. Accurate prediction of ammonia concentration is critical for optimizing treatment processes and ensuring regulatory compliance. This study presents an approach for integrating temporal deep learning models, LSTM and GRU, with explainable feature selection techniques to improve the accuracy and transparency of ammonia concentration prediction in wastewater systems. It also evaluates several machine and deep learning models on a novel dataset collected from North Water’s sewage facility in Delfzijl, Netherlands (2014–2017). In parallel, we apply and compare multiple feature selection techniques, including Backward Elimination, Recursive Feature Elimination, and Sequential Feature Selection, to identify the most informative predictors. These techniques are leveraged not only for model optimization but also as part of an Explainable AI framework to enhance model interpretability and transparency. To further validate the significance of the selected features, we employ SHapley Additive exPlanations to quantify each feature’s contribution to model predictions. The experiments show that the LSTM model combined with Recursive Feature Elimination achieved the best performance (R 2 = 0.96, MAE = 1.044, RMSE = 1.49), effectively capturing temporal trends and highlighting key predictors such as influent nitrite and dissolved oxygen levels.