A process-aware management information system utilizing HRT-aligned machine learning for predicting energy consumption and effluent BOD₅ in WWTPs

Reliable prediction of biological treatment performance in Wastewater Treatment Plants (WWTPs) necessitates data structures that maintain the physical causality between influent characteristics, operational actions, and effluent quality. In this study, we propose a process-aware machine-learning framework that reconstructs operational and laboratory variables based on the mean hydraulic retention time (HRT) trajectory of each influent composite sample. Leveraging detailed Supervisory Control and Data Acquisition (SCADA) measurements, equipment status logs, and laboratory analyses from the Metamorfosis Wastewater Treatment Plant (MWWTP) in Athens-Greece, two process consistent datasets were created. Unlike conventional machine-learning approaches that rely on calendar-based aggregation of operational and laboratory measurements, the proposed framework reconstructs process-residence trajectories via hydraulic retention time, thereby preserving the causal link between influent conditions, operational actions, and treatment outcomes.Dataset 1, spanning 31 August 2024 to 1 December, 2025, captures the energy consumption of the aeration, mixing and Return Activated Sludge (RAS) system. Dataset 2, covering 31 August 2024 to 31 August, 2025, represents hydraulically redistributed effluent Biochemical Oxygen Demand over five days (BOD₅) concentrations. Operational features were engineered exclusively within the aeration residence interval, ensuring temporal alignment with the biological treatment phase and a Random Forest regression model was implemented. The results demonstrated excellent predictive performance for Dataset 1, showing a strong correlation between predicted and observed energy demand due to the direct mechanistic linkage between operational predictors and electrical load. For Dataset 2, the model achieved moderate but robust predictive accuracy. The slightly reduced performance is attributed BOD₅ removal occuring across both aeration and secondary sedimentation phases, while operational variables represent the biological reactor's time window. Taken together, these results demonstrate that hydraulic-time alignment and physically grounded feature engineering significantly enhance model interpretability, generalization, and operational relevance. The principal novelty of the proposed framework lies in integrating hydraulic-process knowledge directly into dataset construction, enabling machine-learning models to learn from causally consistent observations rather than conventional calendar-aligned records. This process-aware methodology provides a scalable and transferable paradigm for developing physically interpretable, operationally meaningful data-driven models in wastewater treatment systems.