Machine Learning-Based Early Warning Framework for Algal Bloom Dynamics in Eutrophic Drinking Water Reservoirs under Thermal Stratification

Harmful algal blooms (HABs) in eutrophic drinking water reservoirs pose severe ecological threats and operational challenges for water utilities, particularly under climate-induced thermal stratification. Traditional monitoring approaches often lack the predictive capability required to implement timely mitigation measures. This study presents a machine learning-based early warning framework designed to forecast cyanobacterial bloom dynamics by integrating high-frequency physical water column data with chemical and meteorological parameters. Using a three-year continuous dataset from a deep, eutrophic drinking water reservoir, dynamic thermal stratification metrics—including the Schmidt Stability Index (SSI) and vertical temperature gradients—were engineered alongside physicochemical features. A hybrid eXtreme Gradient Boosting and Long Short-Term Memory (XGB-LSTM) architecture was developed to predict chlorophyll-a (Chl-a) and phycocyanin (PC) concentrations across 3-day, 5-day, and 7-day lead times. The proposed framework achieved high predictive accuracy ($R^2 = 0.89$, $\text{RMSE} = 3.42\,\mu\text{g/L}$ for 3-day Chl-a forecasts), significantly outperforming standalone machine learning baselines. Feature importance analysis demonstrated that thermal stability indices significantly enhanced early-stage bloom predictability, providing critical lead time for operational responses such as selective intake switching and algaecide dosing. This study provides a scalable, data-driven approach for proactive drinking water safety management under changing thermal regimes.