Weekly Cyanobacterial Bloom-Risk Early Warning Using Deep LSTM and Hybrid CNN-LSTM Models in Jinyang Lake, South Korea

Cyanobacterial blooms threaten the water supply in regulated reservoirs, but the weekly early-warning skill and the usefulness of basin-scale water-quality gradients are not-well quantified. This study created a weekly bloom-risk framework for Jinyang Lake, South Korea, based on the probability of cyanobacteria exceeding 1000 cells/mL in the next 1–4 weeks. Weekly data from 16 stations (2011–2025) trained Deep Long Short-Term Memory (LSTM) and hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) classifiers (training 2011–2022, tuning 2023, independent testing 2024–2025; 48 weeks). The results of principal component analysis (PCA) indicated that the first two components accounted for 46.9–88.1% of the variance by variable, so no single station represented basin-wide conditions. At the 4-week horizon, the hybrid CNN-LSTM achieved F1 = 0.849 (95% CI: 0.74–0.93) and critical success index (CSI) = 0.738 at Naedong, and F1 = 0.758 (0.62–0.88) and CSI = 0.610 at Panmun, exceeding Random Forest (F1 = 0.58–0.62) and XGBoost (F1 = 0.55–0.58). At Naedong, the area under the curve (AUC) was 0.523, while, at Panmun, it was 0.768, due to the class-weighted training that resulted in probabilities being squeezed into a narrow range. PCA-derived features numerically increased the F1 score at Panmun by 0.020–0.041 (not statistically significant) and reduced the Naedong hybrid F1 score from 0.849 to 0.786. Platt scaling reduced hybrid-model Brier scores by 27–29% at both lake stations, improving the usability of the probabilities for operational decisions. The multi-week bloom-risk framework offers reservoir operators and municipal drinking-water systems probabilistic early-warning decision support for preventive action before critical bloom development.