Abstract ID: 3.10717 | Accepted as Poster | Poster | TBA | TBA

Long-term streamflow data is crucial for managing water resources, providing insights into water availability and variability over time. It supports irrigation, drinking water supply, and industrial needs while helping predict and mitigate extreme events like floods and droughts. Additionally, streamflow data is essential for assessing climate change impacts on hydrological systems, aiding policymakers in developing adaptive strategies. However, continuous streamflow measurement in mountainous regions like the southern slope of the Tibetan Plateau is highly challenging due to rugged terrain, extreme weather, and remote locations. Steep slopes and high altitudes make installing and maintaining monitoring equipment costly and difficult. Seasonal variations, such as monsoonal floods and snowmelt, cause rapid discharge changes, complicating measurements. Limited accessibility and resource constraints further hinder the establishment of reliable hydrological monitoring networks. To address these challenges, innovative approaches like satellite-based observations and data-driven models are needed to complement ground-based measurements. Developing physically based models requires extensive spatial data, computational resources, and a deep understanding of hydrological processes, making them complex and resource-intensive. In contrast, data-driven models like Transformers offer a scalable solution for simulating streamflow in data-scarce regions. This study leverages ERA5 precipitation data and available observed streamflow records to develop a Transformer-based model for streamflow simulation from 1940 to 2023. The model is initially trained in data-rich catchments to ensure accuracy before being transferred to sub-catchments with limited data. This transferability approach allows the model to utilize shared hydrological characteristics and forcing data, making reliable predictions in poorly monitored regions. By enhancing the applicability of streamflow models across diverse basins, this method ensures hydrological predictions are accessible for effective water resource management in data-scarce areas.