An Online Adaptive Intelligent Dosing Framework for Drift-Aware Industrial Fluoride Removal

Calcium-salt precipitation is a widely used compliance-control process for industrial fluoride-containing wastewater. Its core objective is to maintain stable effluent fluoride compliance while avoiding unnecessary chemical consumption. In real industrial defluorination systems, however, effluent fluoride is not an instantaneous response to the current influent quality and dosing action. The process therefore exhibits hydraulic delay, nonlinear response, uneven sampling, and process drift, which make a single fixed offline model unreliable for long-term prediction and dosing decisions. Hence, this paper proposes an online adaptive intelligent dosing framework for industrial fluoride removal. First, influent-side variables, dosing actions, and effluent fluoride measurements are aligned according to hydraulic retention time, and lagged variables, rolling statistics, exponentially weighted memory, and first-order differences are constructed as time-enhanced tabular features. Second, a monotonic XGBoost model is adopted as the backbone predictor. The prior that increasing calcium dosage should not increase predicted effluent fluoride is embedded as a monotonic constraint, yielding a more physically consistent dosage-response model. Third, online dosing is formulated as a penalty-based dosage optimization problem to generate a base recommended dosage, and an adaptive safety dose band constrains the executable dosage to reduce underdosing and overdosing risks. Experiments show that time-enhanced features improve short-term prediction under similar operating conditions, the monotonic backbone provides reasonable dosage-response trends, and historical replay together with drift simulations verifies the effectiveness of local adaptation and safety constraints. The proposed method provides a practical solution for intelligent calcium dosing in industrial fluoride wastewater treatment by balancing prediction rationality, online adaptability, and deployability.