How it's made: How Scalable Downscaling Contributes to Your Weather Forecast
In the How It’s Made series is a 20-part series that explains what takes your weather forecast from good to great. We don’t just tell you that our weather forecasts are accurate, we show you how it’s done.
Following on from our post on Model Output Statistics (MOS), we’re continuing our exploration of Category Three: Statistical Post Processing, with this post on scalable downscaling.
Make sure you don’t miss out on the rest of the series. Sign up for the weekly How It’s Made Update. Stay informed and get weekly updates, delivered straight to your inbox.
How can you arrive at a reliable forecast for a specific site, regardless of where it is in the world? In short, the answer is ‘downscaling’, which means weather experts can produce a forecast for every location, based on observations and forecasts from surrounding sites.
The weather model data the experts use for forecasting can cover large grid areas of up to 50km. As a result, it doesn’t always consider the impact of features that fall within those grid lines, such as changes in elevation and land-use. Downscaling helps to overcome this problem—during the process of downscaling, data from the large grid areas is enhanced with geostatistical methods to make predictions about weather and climate ‘on-the-spot’, particularly at the surface level.
The Scalable Downscaling (ScaDo) forecasting system combines the quality of station-based MOS with information about land topography and use, which makes it particularly effective in locations without any observation stations nearby.
"Several weather parameters, like temperature and wind, are highly dependent on altitude and land-use. We are happy we are able to provide a good forecast for every location using ScaDo as a smart and flexible downscaling tool"
- Wim van den Berg
Senior Meteorological Scientist
Where is scalable downscaling used?
To downscale MOS forecasts, the weather experts use data from neighboring MOS stations or stations that are similar in altitude. Where the density of MOS stations is high, and the terrain is not too complex (e.g. no intricate coastlines, deep valleys or steep mountains) that works fine. However, in more complex terrains and areas with just a few MOS stations, the experts will start using ScaDo, which combines the quality of station-based MOS with all kinds of terrain effects like altitude, slope, land-use, nearby water, city conditions, etc.
By combining the quality of station-based MOS with information about land topography and use, ScaDo forecasting systems become highly effective in locations that don’t have observation stations.
How does ScaDo work?
Let’s take an example. Say you require a forecast for an Alpine mountaintop that doesn’t have a weather station. To arrive at a reliable weather forecast for that specific site, the experts look at all the other weather stations in the region. Those observation posts are located at various elevations and certain distances from the site in question, so they have to assign all of these observations a certain weight. This process is called interpolation.
For our Alpine mountaintop, for example, you would receive an algorithm that forecasts the maximum temperature by assigning weights to the surrounding points: A (30%), B (20%), and C (50%). If our site is at an elevation of 3,000 meters, and the surrounding points are located at least 500 meters lower, then we will also have to apply general meteorological algorithms to arrive at a reliable forecast. The innovation primarily lies in the increased flexibility of the algorithms, because ScaDo takes better account of the continually changing, specific weather conditions in the region. Because ScaDo is trained for the whole region, it can also provide a forecast for a location without nearby weather stations.
In mountainous areas, standard downscaling uses a static altitude corrector for temperature, wind and precipitation levels. In ScaDo, the altitude corrector is not static but dynamic; based on raw ECMWF data and MOS station data it takes into account actual weather conditions.
Choosing between numerical modeling and statistical processing
When it must be decided whether to rely on numerical modeling or statistical processing, the weather experts will consider the customer’s requirements and choose the method that will best suit their needs. Typical considerations include:
● (Geo-)statistical downscaling (like ScaDo) delivers a value on any location and therefore requires the kind of data points supplied by observations or MOS for different landscapes, land use, and altitudes. Numerical models can be deployed anywhere, even
when observations are unavailable, but a numerical model always results in average forecasts for grid boxes with “grid-averaged” terrain and land-use and not in a point forecast. A hybrid approach is also possible, i.e., geo-statistical downscaling on numerical model data points; however, this only works when model data is available for different landscapes, land use, and altitudes. While this hybrid approach is less expensive than running a microscale numerical model, it lacks the advantage of starting from unbiased data points like MOS.
● MOS, by nature, scores better on point verification metrics like the mean absolute error.
● As a numerical model uses physics, the output can be more extreme than the input (due to factors such as funneling, katabatic wind effects or focusing of waves due to refraction). The advantage is that extreme events are better predicted.
● Each form of processing needs to be analyzed for its fitness for purpose and the cost versus quality of the output
● When a customer requires a detailed hindcast study, the high-resolution numerical model can also be used to run a forecast in the past
As ScaDo uses Multimodel MOS data for training, it benefits from information from a variety of weather models. Near MOS stations, the accuracy of the ScaDo forecast almost matches the high-level of accuracy achieved with MOS.
Compared to a numerical model ScaDo provides the following unique data types:
● Expert derived elements like, effective cloudiness and weather type
● Forecast along a route e.g. along power lines or a road
● Microscale gridded forecasts that can be applied across the globe
Using the cloud to scale with ease
ScaDo is also cloud-based, which means it doesn’t rely on in-house or on-site computing power to process large volumes of data. In fact, thanks to the cloud, ScaDo can virtually handle unlimited amounts of data cost-effectively, making it particularly useful for projects where budget and location constraints need to be taken into account. Developments never come to an end; the experts are always on the lookout for new possibilities and opportunities to help customers get the weather data they need.
Sign up for the weekly How It’s Made Update. Stay informed and get weekly updates, delivered straight to your inbox.
Several weather parameters, like temperature and wind, are highly dependent on altitude and land-use. We are happy we are able to provide a good forecast for every location using ScaDo as a smart and flexible downscaling tool