Interpretable forecasting of dissolved oxygen leveraging foundation model for proactive aeration in rural wastewater...

Accurate forecasting of dissolved oxygen (DO) is crucial for maintaining biological treatment and minimizing energy consumption in wastewater systems. This challenge is acute in rural stabilization ponds, where DO responds nonlinearly to nutrient fluctuations, episodic inflows, and seasonal variation. To address this, an interpretable forecasting framework was developed using a transformer-based foundation model for time series, fine-tuned on nearly one year of multivariate sensor data from a rural wastewater facility. Four seasonal models (spring, summer, fall, and winter) were trained and evaluated at a 24-hour horizon, corresponding to daily operational planning, with ablation experiments conducted at 1 hour and 168 h to assess robustness. Benchmark comparisons against Support Vector Regression (SVR), XGBoost, Long Short-Term Memory (LSTM), and the transformer-based Temporal Fusion Transformer (TFT) demonstrated substantial improvements. At 24 h, the symmetric mean absolute percentage error (SMAPE) fell from 38 to 46 % for classical machine learning baselines (SVR, XGBoost) and 16-25 % for deep learning benchmarks (LSTM, TFT) to below 7 % with the proposed framework, representing significant gains in accuracy and stability. Interpretability was systematically integrated: SHapley Additive exPlanations (SHAP) attributions identified pH, conductivity, temperature, turbidity, and ammonium as regime-specific drivers, while sensitivity analyses enabled actionable "what-if" exploration. These findings highlight the importance of integrating foundation modeling, seasonal segmentation, and SHAP-based interpretability in improving forecasting for data-constrained rural wastewater systems. The work provides a framework for transparent and decision-aligned DO prediction, with potential relevance for proactive aeration planning and sustainable operation of decentralized wastewater treatment systems.