Adaptive Frequency-Temporal Networks for Climate-Resilient Reservoir Prediction

Accurate forecasting of reservoir water volume is crucial for ensuring regional water security amid intensifying climate change and increasing anthropogenic disturbances. Traditional hydrological models often struggle to capture the nonlinear and multiscale dynamics of reservoir systems shaped by both natural variability and human interventions. In this study, we propose a novel hybrid deep learning framework that integrates FilterNet and TimeMixer to improve the accuracy and robustness of reservoir water volume prediction. A multi-scale forecasting architecture is designed to simultaneously capture short-term fluctuations and long-term climatic trends by decomposing input sequences into distinct temporal resolutions and modeling them with decomposition-aware temporal-channel mixing. We explicitly incorporate human regulation data - such as water diversion and operational scheduling - into the forecasting pipeline, enhancing the model's practical adaptability to real-world reservoir operations. Adaptive frequency-domain filtering techniques are employed to mitigate noise from sensor errors and data inconsistencies, thereby improving signal stability and temporal feature extraction. Experimental results on a real-world reservoir dataset demonstrate the model's superior performance over existing baselines, highlighting its effectiveness in capturing both short-term fluctuations and long-term trends. This work contributes a generalizable approach for coupling climatic and anthropogenic signals in hydrological forecasting, with significant implications for intelligent reservoir operation and adaptive water resource management.