Reservoir Storage Forecasting Using Machine Learning in the Troy Region, Türkiye

Accurate prediction of dam reservoir storage is essential for sustainable water resource management, irrigation planning, and drought preparedness. In this study, the monthly storage volumes of the Atikhisar and Bakacak reservoirs in the Troy Region of Northwestern Türkiye were predicted using meteorological variables, the Standardized Reservoir Index (SRI), and reservoir memory variables. Four prediction targets were considered: current-month storage V(t), one-month-ahead storage V(t + 1), current-month storage change ΔV(t), and one-month-ahead storage change ΔV(t + 1). Eight machine learning models were evaluated using time-series cross-validation and out-of-fold prediction metrics. The results showed that meteorological variables alone were limited in explaining absolute reservoir storage, especially for the Bakacak Reservoir, whereas they provided more useful information for storage change targets. The inclusion of reservoir memory variables substantially improved the prediction performance, highlighting the dominant role of antecedent storage conditions. The highest accuracy was obtained using the integrated model combining meteorological variables, the SRI, and reservoir memory. The final models achieved OOF R2 values of up to 0.967 for Atikhisar and 0.980 for Bakacak. SHAP analysis indicated that the SRI, antecedent storage, storage occupancy, and lagged meteorological variables were the main drivers of the model predictions. Overall, the proposed framework provides an interpretable and practical decision support tool for short-term reservoir operation, irrigation planning, and drought-sensitive water resource management.