Machine learning-driven prediction and control system for practical application of sulfur-based autotrophic denitrification biofilters

Sulfur-based autotrophic denitrification (SADeN) offers a sustainable, low-carbon alternative for advanced nitrogen removal, yet its efficacy is highly susceptible to fluctuating environmental conditions. Conventional control strategies often fail to address these complex, nonlinear interactions, leading to unstable performance and chemical overuse. Hence, we developed the SADeN prediction and control system (SADeN-PaCS) for cross-season prediction, interpretable diagnosis, and adaptive regulation of full-scale SADeN biofilters. The artificial neural network (ANN) achieved high accuracy in predicting effluent nitrate (NO₃⁻-N_eff, R² = 0.94). Model interpretability revealed the in-situ kinetics and state-dependent dynamics, identifying water temperature (WT) as the primary determinant of system resilience and specific filler nitrate loading (SFNL) as the key controllable variable. Interaction analyses further demonstrated that low WT and elevated dissolved oxygen within the media layer (DO_m) imposed coupled inhibition by suppressing sulfur-autotrophic metabolism and disrupting the denitrifying microenvironment. These findings informed a dual-mode regulation strategy: under the benefit-maximization mode (WT > 18 °C), SFNL was increased by 36.07% (from 0.61 to 0.83 kg-N/m³/d) to enhance treatment capacity; under the risk-mitigation mode (WT ≤ 18 °C), SFNL was proactively reduced to ensure effluent quality. For extreme scenarios (low-WT and high-SFNL), an additional chemical dosing model (R² = 0.91) was incorporated to optimize thiosulfate supplementation, enabling stabilized NO₃⁻-N_eff below 5.0 mg/L while reducing chemical use by 59.90% versus empirical approaches. Overall, this work pioneers a diagnostic-driven control paradigm for SADeN process management, leveraging model interpretability to translate predictive insights into adaptive, multi-level operational strategies for enhanced resilience and cost-efficiency.