Beyond data-driven models: A dedicated temporal odorous risk classifier for varying-temperature scenarios based on dual odor-producing mechanisms.

Odorants are common reasons for complaints on drinking water. Temporal odorous risk prediction is thus essential but remains unsolved. Unlike static temperature, varying temperature lacks dedicated straightforward estimating methods for the corresponding odorous risks and the existing ones are data-driven without odorous knowledge. This study developed a 4-layer hybrid model to progressively classify risks with dual odor-producing mechanisms (external: exposure-lag-response relationships; internal: precursor supply-oxidative stress relationships), using multi-day temperature sequences. Daily total odorant production (TOP), cell intensity (CI), and cell quota (CQ) were measured. Distributed lag nonlinear models (DLNM) revealed that thermal effects were lag-dependent: promotion at short lags (0-2 d), whereas inhibition at longer lags (3-7 d). LightGBM learned exposure-lag-response patterns with test R2 of 0.837-0.948. A dual-teacher data-mechanism classifier discriminated three risk levels with AUC of 0.901-0.965. Independent experiments on new sequences reached 87.5% accuracy. Untargeted metabolomics indicated that high risk emerged when precursor supply, cleavage activity, and growth support were coordinated within the same sequence, rather than when absolute temperature was high alone. By introducing internal metabolic evidence into a pathway-informed Bayesian hybrid model (PIBHM), balanced accuracy was improved by 3.09-3.22%. Since weather forecasts are freely available, this work offers a practical warning tool, validated under laboratory conditions, for estimating odorous risks in water treatment operations before odorants are perceptible.