Entropy-Weighted TOPSIS Fusion Ensemble for Robust Dissolved Oxygen Soft Sensing with Interpretable Decision-Support Insights in a Full-Scale Wastewater Treatment Plant

To address the challenges associated with elevated salinity-related background conditions and low temperature in wastewater treatment plants (WWTPs) in Northwest China, this study investigates dissolved oxygen (DO) soft-sensing and interpretable decision support for under-instrumented conditions. Existing approaches often depend on online chemical instrumentation, whose maintenance burden and unstable long-term operation limit their applicability in resource-constrained plants. To address this gap, a DO soft-sensing and interpretable decision-support framework was developed and evaluated in a full-scale plug-flow Anaerobic-Anoxic-Oxic (AAO) WWTP in Northwest China. Using 15,676 hourly observations collected across seasons, an Entropy-Weighted TOPSIS Fusion Model (ETFM) ensemble was constructed through Bayesian hyperparameter optimization and multi-model benchmarking to improve prediction stability under complex operating regimes. Under the random-split benchmark, ETFM achieved an R 2 of 0.966 while additional time-ordered validation retained acceptable temporal generalization with an R 2 of 0.915. Multi-level interpretability analyses based on SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP), 3D Partial Dependence Plots (3DPDP), and 3D Accumulated Local Effects (3DALE) were then used to characterize the nonlinear relationships learned by the model. These analyses suggested model-inferred diminishing marginal returns for upstream DO, beyond which further increases in upstream DO provided limited predicted downstream benefit. Based on these model-explanation results, a conceptual decision-support logic was proposed to support state-informed downstream DO soft sensing, ORP-based process-state screening, and preliminary decision-support evaluation.