Incorporating multi-timescale data into a single long short-term memory network to enhance reservoir-regulated streamflow simulation

Water security and its sustainable management are critical to human survival and livelihoods. Under the pressures of climate change and population growth, an increasing number of natural watersheds are being regulated by dams and reservoirs. However, the operation of these man-made infrastructures, particularly small-scale reservoirs managed by local governments, is often characterized by flexible and irregular management practices, significantly complicating streamflow modeling-a critical aspect of water management. Remote sensing provides valuable reservoir storage data for data-scarce basins, but its coarse temporal resolution requires integration with ground-based observations or simulations. Unlike traditional physical models that rely on explicit hydrological processes and predefined reservoir operation rules, or data-driven methods that struggle with multi-timescale data and their dependencies, the Multi-TimeScale Long Short-Term Memory (MTS-LSTM) model is a deep learning framework designed to integrate multi-timescale data. This study evaluates the MTSLSTM in integrating monthly remote sensing-derived reservoir storage variation data to simulate daily reservoir-regulated streamflow. The case study on the Yuanjiang River Basin demonstrated that the MTS-LSTM effectively bridges the gap between SWAT-simulated natural streamflow and observed regulated streamflow, a gap primarily caused by reservoir storage variations. The model achieved strong performance at two hydrological stations. For monthly simulations, the mean Correlation Coefficient (CC) was 0.92, Nash-Sutcliffe Efficiency (NSE) was 0.81, and Kling-Gupta Efficiency (KGE) was 0.80. For daily simulations, the mean CC was 0.79, NSE was 0.58, and KGE was 0.71. Integrating remote sensing data significantly enhances simulation accuracy, outperforming naive LSTM models. This study presents a systematic methodology for incorporating multi-source and multi-timescale data to enhance the accuracy of reservoir-regulated streamflow simulations, with a particular focus on regions with limited data and hybrid cascade reservoir systems.