Science:Downscaling

From SwissExperiment

Contents 1 What is Downscaling? 2 Statistical Tools 2.1 Goals 3 Downscaling Wind Data 4 References

What is Downscaling?

Downscaling is a catch-all expression for the transformation of data collected at one point in space (or time) to another, using a combination of statistical tools (regression, filtering and so-on) and physical models (meteorological models, etc.). The techniques are honed using previous data, which means that downscaling allows gaps in data coverage to be filled with a known - albeit historic - margin of error.

Statistical Tools

Several tools have been implemented within the wiki to help in the downscaling process. They include regression, data analysis, smoothing tools and neural networks; an overview is given below.

Tool	Input data	Output data	Transfer function	notes
x-y plot	pairs of x-y data	-	-	Template:ScienceRxyplot
x-y linear regression	pairs of x-y data	-	-	Simple linear regression using LSE. See Template:ScienceRxylinreg
Holt-Winters filter	pairs of x-y data	-	-	Adaptive filtering using the holt-winters filter. See Template:ScienceRHoltWintersfilter

These tools have been implemented using the R extension - see In-application data access.

N.B. In principle, all R packages are available, but they do have to be installed on the server by the wiki administrator. Please contact them for details; a list of existing R packages is given in In-application data access.

Goals

The goal of these tools is that we can implement a very simple workflow within the wiki;

1. get data from GSN for a particular site for a particular time period
2. perform some operation on it using R
3. display the results

It might be possible to implement this as a form or query, so that a user can select (for example)

• forecast data source (e.g. NWP data, implemented as a GSN virtual instrument)
• reference location (again, a GSN virtual instrument)
• forecast variable (e.g. temperature or wind speed or...)
• forecast time period

The template used would then

• import recent historic NWP data and forecast
• perform smoothing to 'clean' the reference location data
• perform correlation or neural network training using the existing data from the NWP and reference location
• forecast the reference site conditions from the correlation and forecast NWP data
• report the results back to the wiki

Downscaling Wind Data

A recent study by Meteotest and MeteoSchweiz ([Dierer, 2009, p. 25]) identified several wind speed predictors that could be retrieved from NWP and used to predict wind speed FF at other points.

Predictor	Description	Source
FF(10)	wind speed at 10 m above ground	NWP
FF(100)	wind speed at 100 m above ground [m/s]	NWP
RH(2)	relative humidity at 2 m above ground [%]	NWP
RH(3000)	relative humidity at 3,000 m above ground [%]	NWP
TT(2)	Temperature at 2 m above ground [C]	NWP
TGRAD	temperature gradient between 3,000 m and the surface [K/m]	NWP
DPWE(3000)	pressure gradient W-E at 3,000 m a.s.l. [Pa/m]	NWP
DPNS(3000)	pressure gradient N-S at 3,000 m a.s.l. [Pa/m]	NWP
FGEO(3000)	Geostrophic wind at 3,000 m a.s.l. [Pa/m]	NWP

The geostrophic wind at a height z, denoted FGEO_(z), is calculated from the NWP data as

• 
• 
• 
• 
• 
• 

A linear regression is applied using the predictors to the observed hourly data at a nearby measurement station, taking the form

The 4 most important predictors are then identified for each of the following times of day...

• day time (0800-1600 UTC)
• night time (200-0400 UTC)
• Transitional periods (0500-0700, 1700-1900 UTC, linear interpolation between daytime and nighttime data)

...and for different times of year

• summer (June, July, August)
• winter (November, December, January, February)
• transitional periods (linear interpolation between summer and winter, e.g. March is 2/3 winter, 1/3 summer)

This linear regression gives two times series for each observation point;

• an observed wind speed FF_obs and
• a modelled wind speed FF_mod

A kalman filter using the modelled and observed wind speeds is then used to estimate the value at the next time step.

References

• Silke Dierer, Christophe Hug, Jan Remund, Beat Schaffner and Vanessa Stauch, 2009 Fore- und Nowcasting der Stromproduktion von Windenergieanlagen in komplexem Gelände. Available online via Swiss Bundesamt fur Energie at http://www.bfe.admin.ch/dokumentation/energieforschung/index.html?lang=en&project=102399.
• Andrew Kusiak, Wenyan Li, 2010. Estimation of wind speed: A data-driven approach, Journal of Wind Engineering and Industrial Aerodynamics, In Press, Corrected Proof, Available online 4 June 2010, ISSN 0167-6105, DOI: 10.1016/j.jweia.2010.04.010.

[[1]]