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

Accurate forecasting of phytoplankton dynamics at drinking-water intake points is essential for proactive reservoir management because sudden biomass increases can directly impair filtration and increase backwashing demand. Using daily monitoring data collected at the Hoedong Reservoir intake depth from November 2012 to December 2021, we developed and evaluated a physics-informed Transformer (PITF) framework to predict the intake-point cell density of two operationally problematic genera (Microcystis and Stephanodiscus). Missing observations were addressed using a centered rolling-median interpolation, and a noise-perturbed dataset was additionally generated to assess robustness under realistic input uncertainty. All models performed were evaluated using R² and RMSE. Intake point monitoring indicated risk patterns, with Microcystis peaking during mid-summer to autumn (up to 43,560 cells·mL^-1) and Stephanodiscus peaking in winter (up to 5,120 cells·mL^-1). PITF achieved strong validation performance for Microcystis (R²=0.711) and reasonable performance for Stephanodiscus (R²=0.505). Under noise-perturbed inputs, PITF showed slight performance degradation (e.g., Microcystis R²=0.701; RMSE=0.761), indicating robust forecasting skill under input uncertainty. A SHAP-based plausibility assessment identified ecologically consistent drivers, with water temperature and dissolved oxygen as dominant predictors for both taxa. Overall, PITF provides an interpretable and robust intake-focused forecasting approach that can support early-warning decision-making for drinking-water treatment operations.