Sub-Seasonal to Seasonal Reservoir Storage Forecasting Using an Attention-Based Temporal Fusion Transformer

Reservoirs play a vital role in water resources management by supporting irrigation, hydropower generation, urban water supply, flood mitigation, drought preparedness, and food security, particularly in monsoon-dominated regions. However, accurate forecasting of reservoir storage remains challenging due to the combined influences of climate variability and anthropogenic regulation, which limit the reliability of traditional hydrological and statistical models. In this study, we systematically evaluate four deep learning approaches Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional LSTM (CNN–LSTM), and the Temporal Fusion Transformer (TFT) for reservoir storage prediction across 79 major reservoirs in India. The models are trained using multivariate inputs that comprise historical storage, precipitation, and temperature data spanning the years 2000–2023. Our results demonstrate that the TFT consistently outperforms the recurrent and convolutional baselines, achieving testing coefficient of determination (R²) values exceeding 0.95 across most reservoirs and reducing prediction errors by approximately 20–30% relative to LSTM- and GRU-based models. Building on this superior performance, we conduct sub-seasonal to seasonal forecasts with lead times of up to three months. The TFT exhibits strong drought detection capability, with Probability of Detection values exceeding 0.8 for 1–3-month lead times. Furthermore, purely data-driven TFT forecasts outperform simulations that incorporate external precipitation and temperature forecasts from the Climate Forecast System Version 2 (CFSv2), highlighting the robustness of the learned temporal representations. Overall, this study demonstrates the potential of transformer-based deep learning models to enhance reservoir storage forecasting and early warning capabilities, offering a promising pathway for improving adaptive reservoir operations and water resources management under hydroclimatic variability.