Dataset for "An operational river-stage forecasting system using spatiotemporal graph neural network (ST-GNN) in Ascension Parish, Louisiana"

Abstract

This dataset accompanies the manuscript "An operational river-stage forecasting system using spatiotemporal graph neural network (ST-GNN) in Ascension Parish, Louisiana", prepared for submission to Hydrology and Earth System Sciences (HESS). It provides the evaluation data for a 68-gauge forecasting network comprising 51 scored in-parish gauges and 17 supporting boundary gauges. The paired forecast and observed stage series cover 5,832 hourly forecast origins from 1 January to 31 August 2026, with 96 lead steps at 15-minute intervals over a 24-hour horizon. The evaluated ST-GNN, GRU, and LSTM use final weights fitted to eligible data through 31 December 2025 and frozen during the 2026 evaluation, with three training seeds (101, 202, and 303). Model columns contain the P80 forecasts scored in the study; persistence is included as a benchmark. The event-only subset comprises 927 origins when at least one parish pump was running at issue time. Three compressed CSV files provide the observed-rainfall hindcast, the retrospective simulated operational forecast using issue-time HRRR rainfall and operational postprocessing, and the ST-GNN rainfall-forcing comparison without postprocessing. The deposit also includes the gauge inventory; graph definitions for final fitting, chronological epoch selection, and gauge-network sensitivity; per-gauge evaluation metrics; full-precision inputs for paired bootstrap uncertainty analysis; archived rainfall forcings; and figure inputs. Directory contents are packaged in the corresponding RAR archives. README.md describes file schemas, scoring and aggregation rules, and the mapping to the manuscript results. FILE_MANIFEST.csv and public_source_manifest.csv document file integrity and public input sources. Stages are expressed in metres and the paired CSV values are rounded to 0.1 mm. Observed stages are masked for missing or stale gauge readings under the documented screening rule. The deposit supports evaluation reproducibility; the complete feature-engineered training matrix is not included. Internal Ascension Parish Government HEC-RAS terrain, simulated water-surface, and mesh files are not redistributed, although their derived graphs are included. Public hydrologic and meteorological inputs are documented by retrieval source; rainfall forecasts are included as used. The companion code and trained-weight record is on github https://github.com/awesomemfg/River_ST_GNN_Forecast. It contains the data preparation, graph construction, chronological training, inference, postprocessing, and evaluation workflow. This data record is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).