Although the native temporal resolution is weeks, months, and quarters, Salient offers a downscaling feature to create a realistic time series from day 1 to day 365 at a daily resolution for numerous variables.
Salient uses forecast anomalies for temperature, precipitation, or winds speed to select analogs, or past years with the most similar weather conditions. Analog selection depends on the variable that is being requested. For example, we use 10m wind speed to select analogs related to wind speed.
The hourly data is selected from the historical analogs for the daily timeseries.
Hourly Downscaling is available for the following variables: temp, precip, rh, wspd, wdir, wspd100, wdir100, wgst, cc, tsi, dni, dhi, snow, sm, st
Hourly downscale is available for the following variables: Temperature (2m), Precipitation, Relative Humidity, Wind Speed (10m, 100m), Wind Direction (10m, 100m), Maximum Wind Gust (10m), Cloud Cover, Total solar insolation, Direct Normal Irradiation, Diffuse Horizontal Irradiation, Snow Cover, Soil Moisture, and Soil Temperature
Downscaling is most effective when done in an ‘ensembling’ manner. The use of multiple analogs (creating an ensemble of analogs) allows you to access many different realistic simulations of daily weather, conditioned on Salient's forecast anomalies. You can think of the ensemble of analogs generated by daily downscale, to be the same as the ensemble members generated by numerical weather prediction. The spatiotemporal relationships between variables are preserved in our analog selection process. The ensemble of analogs generally reflects our native resolution forecasts, and provides a set of plausible daily scenarios that could have led to the weekly, monthly, or seasonal forecast anomaly.
Since each ensemble member represents an additional analog, increasing the number of ensemble members increases representativeness. We sample historical analogs from the available probability distribution. The greater the number of ensemble members, the closer it represents the full distribution. The optimal number of ensembles depends on multiple factors such as variable and location, but we recommend a minimum of 50 members.
To run the daily time series through an internal model, call the endpoint with ensembles=50, for example, to download 50 ensemble members. Run each ensemble member through your model and then average your model output. If you average the 50 ensemble members first and then run them through your model, this washes out the variability so that each day in the time series is close to the native temporal resolution.
You can download the downscaled forecast for a whole grid by using the shapefile argument. (Note that the shapefile input disables debiasing). The geospatial file should be no larger than 8GB for performance and should be meteorologically consistent, because historical analogs are consistent for the shapefile in order to ensure spatial consistency**.**
You can also apply downscale to a list of coordinates by providing a csv file of multiple lat/lons. When using a vector of coordinates, the locations should be meteorologically consistent, because historical analogs are consistent across the coordinate list in order to ensure spatial consistency.
See Upload for additional help.
<aside> 💡 Note:
https://www.notion.so/images/favicon.ico
</aside>