Overview
Improve the representation of river dynamics in regional flood-forecasting research.
How it works
The current Hybrid4 configuration averages four members with equal weights in physical discharge space: three event-aware raster Transformers with different random seeds and one persistence-residual histogram gradient-boosted tree model. Inputs combine gauge observations, ERA5-Land reanalysis and spatial rasters. Forecast-horizon ERA5-Land is supplied in this retrospective setting, so the illustration describes reanalysis-assisted evaluation rather than an operational weather-forecast deployment.
Three event-aware raster Transformers (seeds 42, 43 and 44) and one persistence-residual HGB member are converted to physical discharge, clipped and averaged with equal weights.
AI-assisted method illustration
Gauge observations, ERA5 reanalysis and geospatial raster layers provide temporal and spatial context. These are schematic inputs, not exported research data.
AI-assisted method illustration
Historical observations precede a three-step discharge target. ERA5 values in the forecast horizon are available as retrospective reanalysis; operational forecasting would require a corresponding weather-forecast input.
AI-assisted method illustrationSelected contributions
- Developed a forecasting approach that represents river-network relationships.
- Integrated reanalysis, hydrometric observations and geospatial data covering 75 gauges across 15 rivers.
- Implemented and assessed models for hydrological time-series prediction.
Research status
Second author of “A Spatiotemporal Hybrid Framework for Regional Flood Forecasting”; abstract accepted for poster presentation at WRE 2026. Research under Guangdong Major Project of Basic and Applied Basic Research No. 2021B0301030007, “High-Resolution Simulation and Prediction of the Land Surface Environment”.