Physics-Informed Transformer for Genus-Level Phytoplankton Forecasting at Drinking-Water Intake Points

Area of Science:

• Environmental Science
• Water Resource Management
• Machine Learning Applications

Background:

• Phytoplankton blooms at drinking water intakes disrupt filtration and increase operational costs.
• Accurate forecasting is essential for proactive reservoir management and ensuring water quality.

Purpose of the Study:

• To develop and evaluate a physics-informed Transformer (PITF) framework for predicting phytoplankton cell density at drinking water intake points.
• To assess the robustness of the PITF framework under input uncertainty.
• To identify key ecological drivers influencing phytoplankton dynamics.

Main Methods:

• Utilized daily monitoring data from Hoedong Reservoir (2012-2021).
• Developed a physics-informed Transformer (PITF) framework to predict cell density of Microcystis and Stephanodiscus.
• Employed centered rolling-median interpolation for missing data and generated noise-perturbed datasets to test robustness.
• Evaluated model performance using R² and RMSE, and assessed driver importance with SHAP.

Main Results:

• PITF demonstrated strong predictive performance for Microcystis (R²=0.711) and reasonable performance for Stephanodiscus (R²=0.505).
• The framework showed robust forecasting skill even with noise-perturbed input data.
• SHAP analysis identified water temperature and dissolved oxygen as dominant predictors for both phytoplankton genera.

Conclusions:

• The physics-informed Transformer (PITF) framework offers an interpretable and robust approach for forecasting phytoplankton dynamics at intake points.
• This method supports early-warning decision-making for drinking water treatment operations.
• PITF enhances proactive reservoir management by predicting problematic phytoplankton genera.