Advancing Near-Term Water Quality Forecasting with Explainable Machine Learning and Probabilistic Uncertainty Quantification
TL;DR
An end-to-end framework integrating an extreme gradient boosting (XGBoost) ML with automated hyperparameter optimization, bootstrap-ensemble uncertainty quantification, and interpretability through SHapley additive exPlanations (SHAP), which was applied to forecast phytoplankton in a water-supply reservoir.
Abstract
Phytoplankton blooms threaten drinking-water security, yet translating high-frequency monitoring data into timely, reliable, and interpretable forecasts remains challenging. Machine learning (ML) models can improve predictive accuracy but typically require intensive tuning, yield deterministic forecasts without uncertainty, and offer limited model explainability. To address these limitations, we present here an end-to-end framework integrating an extreme gradient boosting (XGBoost) ML with automated hyperparameter optimization, bootstrap-ensemble uncertainty quantification, and interpretability through SHapley additive exPlanations (SHAP), which we applied to forecast phytoplankton in a water-supply reservoir. A 10-member ML ensemble achieved high chlorophyll-a forecast skill and uncertainty quantification, improving continuous ranked probability score (CRPS) by up to 14.6% over the single model, with ensemble CRPS of 1.2 and 2.7 μg/L at 1 and 7 days ahead, respectively. Global SHAP explanations identified initial chlorophyll-a as the primary driver of the forecast 1 day ahead, with thermal stratification and dissolved organic matter gaining prominence at 7 days. Local SHAP attribution showed individual events to be physically interpretable, each traceable to the dynamic influence of thermal stratification, meteorological forcing, and initial chlorophyll-a. This lightweight transferable framework advances mechanistic understanding of phytoplankton dynamics and supports proactive, risk-informed water-quality management.