Improving Deep Learning Streamflow Forecasts by Explicitly Training with Real-world Precipitation Forecasting

Overview This dataset supports the research titled "Improving Deep Learning Streamflow Forecasts by Explicitly Training with Real-world Precipitation Forecasting." It is designed to address the data distribution shift in deep learning-based hydrological modeling caused by discrepancies between reanalysis forcing and real-world Numerical Weather Prediction (NWP). The dataset covers 502 basins, selected and curated for their data integrity and representativeness. Core Components Dynamic Forcing & Streamflow: 6-hourly time series (2016-2023) including MSWEP precipitation, potential evaporation, temperature, snow-related variables, and observed streamflow. GFS Forecasts: Real-world GFS precipitation forecasts (2016-2023) across multiple lead times (6 to 252 hours), specifically formatted for explicit training of deep learning models. Static Attributes & Geospatial Data: Comprehensive physical attributes (topography, soil, climate, etc.) for all 502 basins, supplemented by Shapefiles for basin boundaries and outlets. Evaluation & Classification: Basin-scale GFS performance metrics (RMSE, MAE, CC) and tiered classification lists to facilitate performance analysis under varying forcing uncertainties. Data Structure The data is organized into logical sub-directories (timeseries/, forecasts/, attributes/, shapes/, basin_classify/) with a master ID list (502_ids.csv) and a unit mapping file (units_info.json) to ensure seamless data alignment and integration for researchers.