Hybrid machine learning model for disinfectant dosing in small-scale water treatment under data scarcity

Abstract

Disinfection by-products, including trihalomethanes and haloacetic acids, pose persistent risks to human health and aquatic ecosystems, particularly in small-scale water treatment plants characterized by limited automation and incomplete monitoring records. This study proposes a hybrid model that integrates extreme gradient enhancement with seasonal trend decomposition, allowing incomplete time series to be partitioned into trend and seasonal components, thereby improving prediction stability and improving interpretability of variable influence. The main contribution is a method that explicitly addresses seasonal variability and data scarcity while preserving predictive accuracy under infrastructure constraints, achieving 𝑅2 ≥ 0.90 and RMSE values between 0.15 and 0.30. The model was validated in a real decentralized system, where it exhibited high performance even with data missing up to 30%, producing monthly reductions of approximately 450 g of trihalomethanes and 800 g of haloacetic acids, along with lower chlorine and fluoride consumption. By integrating technical, environmental, and economic dimensions, including measurable financial returns with a positive annual ROI and a short payback period, the approach provides a replicable solution for dosing control in data-limited contexts, aligned with the Sustainable Development Goal 6 of the United Nations and the advancement of responsible digital strategies in the water sector.