An effluent risk informed closed-loop framework for early warning of influent anomalies using COD soft sensing.

Sudden shock loads in wastewater influent can severely disrupt biological treatment processes and cause effluent quality exceedances in wastewater treatment plants, particularly in domestic-industrial integrated facilities. Timely and reliable early warning of such disturbances is critical for enabling proactive intervention and minimizing environmental and operational risks. This study develops a COD-centric closed-loop early warning framework that integrates a machine-learning-based soft-sensing module with a multi-step effluent prediction module, in which anomaly tagging is automatically triggered by effluent discharge limit thresholds. Two coupling architectures were evaluated: a serially coupled architecture (SCA) that sequentially connects influent sensing and effluent prediction, and a jointly coupled architecture (JCA) that enables end-to-end learning within a unified model. Feature importance was interpreted using SHAP analysis and validated through ablation studies to identify the key process variables that govern predictive performance and early-warning responsiveness. In a full-scale integrated wastewater treatment plant (IWTP) case study, the proposed framework achieved 95.0 % accuracy and 87.2 % anomaly detection precision for 12 h ahead warnings, with JCA outperforming SCA. These results demonstrate that backward inference from predicted effluent compliance risk enables timely identification of upstream disturbances before limit violations occur. This integrated and interpretable framework provides a novel, real-time, and cost-effective solution for linking effluent-risk forecasting with influent-anomaly diagnosis, substantially enhancing the operational resilience and proactive management of IWTPs.