Interpretable machine learning for cyanobacterial density prediction and driver identification using sparse and irregular reservoir monitoring data

Drinking water reservoirs face serious threats from cyanobacterial blooms driven by climate change and nutrient inputs from human activities. Traditional statistical methods struggle to handle sparse and irregular monitoring data and the nonlinear relationships between cyanobacteria and environmental factors. However, conventional machine learning models often lack interpretability. This study used water quality and cyanobacterial data collected from 2020 to 2022 in a reservoir in southern China. Akima interpolation was used to reconstruct unsampled daily values. The performance of random forest (RF), extreme gradient boosting (XGBoost), and bidirectional long short-term memory (BiLSTM) models was compared for predicting cyanobacterial density, and key drivers were identified using shapley additive explanation (SHAP). The results indicated that BiLSTM model achieved the highest prediction accuracy, with test R² values of 0.903, 0.849, and 0.919, and corresponding test Root Mean Square Error (RMSE) values of 1.169, 0.460, and 2.160 × 10 6 for Pseudanabaena sp., Cylindrospermopsis raciborskii , and total cyanobacteria, respectively. SHAP analysis revealed different drivers for the three targets. Pseudanabaena sp. was primarily regulated by nitrate nitrogen (NO 3 ⁻-N) and negatively correlated with dissolved oxygen (DO), whereas Cylindrospermopsis raciborskii was primarily driven by dissolved total phosphorus (DTP) and was positively influenced by DO . Total cyanobacterial density was jointly driven by DTP, the permanganate index (COD Mn ), and water temperature (WT). This study provides an interpretable machine learning framework for the cyanobacterial density prediction in drinking water reservoirs and offers guidance for algal bloom early warning strategies.