Abstract
Oxygen transfer in activated sludge varies with wastewater and operating conditions, as reflected by the α₀-factor. We examined how preprocessing affects its prediction and generalization using 18,467 records from four treatment stages at three German wastewater treatment plants. Six machine-learning models were evaluated with preprocessing fitted on training data only and model selection by cross-validation. Feature filtering reduced 17 inputs to 12-15 and changed test R² by at most 0.010. In fixed-setting CatBoost comparisons, 3σ training-row removal lowered test R² by 0.022 to 0.033 within individual datasets; interquartile-range and median-absolute-deviation screening caused larger losses of 0.055 to 0.285. The resulting pipeline retained all training rows and flagged inputs outside training-derived bounds. On random-row test sets, selected models achieved R² of 0.872 to 0.924 within individual datasets, with 95.2 to 99.2% of predictions within ±10%. Blocked holdout reduced R² to 0.005 to 0.556. Chronological holdout produced systematic overprediction and negative R² in A2, B and C, and all evaluated tree models had negative R² in leave-one-dataset-out tests. Thus, random-row accuracy did not establish generalization to later periods or unseen datasets. The findings are limited by unavailable timestamps, dependence between observations, test-informed pipeline selection and the absence of external or operational validation. Feature attributions describe associations rather than causal effects. Independent time-aware validation and site-specific recalibration remain necessary.