Predicting Algal Bloom Dynamics in Drinking Water Reservoirs Using High-Frequency In Situ Data and Machine Learning

Algal proliferation in subtropical drinking water reservoirs has become increasingly severe, and developing a reliable prediction for algal abundance through high-frequency in situ data is essential for early risk warning and effective management. This study analyzed the interannual variations in algal abundance in the Shanmei (SM) Reservoir, located in Quanzhou City, Fujian Province, China, based on the high-frequency data between 2020 and 2025, and forecasted algal abundance 24 h ahead via the optimized Transformer model. Results revealed that the SM reservoir exhibited seasonal variability in environmental factors, with persistently elevated pH during spring and summer, ranging from 7.12 to 9.66, and relatively high total nitrogen concentrations, ranging from 1.17 to 2.28 mg/L. Overall, algal abundance increased throughout the study period, and the annual average algal abundance in 2025 was 8.18 × 10⁶ cells/L, which was twice that in 2021. Model comparisons revealed that the optimized Transformer model exhibited the highest performance in terms of R² = 0.88 when predicting the next hour using 12 days of data. Feature importance analysis based on SHapley Additive exPlanations (SHAPs) revealed that the predictions of algal dynamics were primarily influenced by previous-hours algal abundance, permanganate index, dissolved oxygen, air temperature, wind speed, and pH. This study revealed that the optimized independent learning model with integrated multi-scale features can significantly enhance the predictive performance of algal dynamics, offering a technical basis for early warning of algal blooms and refined reservoir management.