Our Approach

Our models include variables from the world's top weather forecasting models, including:

• ECMWF ERA5-Land Reanalysis

• NOAA's Global Ensemble Forecast System (GEFS)

• NOAA's Climate Forecasting System (CFS)

We use machine learning models which are trained on historical weather data and recordings of river discharge values at specific gauge stations. Additionally, we create gauge-specific rating curves to convert discharge forecasts into river level forecasts.

When training, reanalysis data is used as "forecast" data, giving our models access to how the given river basin has responded to specific weather events in the past. This enables the models to learn the physics of the basin accurately.

When performing inference, we use weather forecasts instead of future weather data since such data does not exist at that point in time. We use the GEFS values from today to their terminal horizon (16 days) and append the CFS forecast beyond that point.

We are able to provide an estimation of the forecast uncertainty using the weather forecast ensembles. The displayed uncertainty bands show the error on the mean forecast derived from all combinations of the 31 GEFS members and four CFS members.

Variables:

• Daily total precipitation

• 2-meter temperature daily mean

• Skin (surface) temperature daily mean

• Percent snow cover daily mean

• Snow depth equivalent daily mean

• Surface latent heat flux daily mean

• Level 1 soil moisture daily mean

• Level 1 soil moisture daily mean