The European forecasting model already outperforms all other global forecasting systems in the world, including the North American GFS model. The most overt demonstration of the European model’s superiority took place the week before Hurricane Sandy’s devastating landfall in 2012. Of more than a dozen computer forecasts, only he showed that the storm would move along a path toward the east coast of the United States. States turned instead of staying. innocent to the sea.
Now the world’s best prediction model keeps getting better, and not just a little bit. An upgrade that went live this week offers dramatic improvements to the model’s resolution, both for deterministic forecasting and for the ensemble model runs used to predict conditions a week or more into the future. “What the European modeling community is doing is just amazing,” Ryan Maue, a meteorologist at WeatherBell, told Ars. “This is the golden age of weather forecasters. It is an absolute marvel of computer modeling technology.”
Maue lives and breathes computer forecasting models every day as he converts the raw output of global forecasting entities, including the meteorological agencies of Europe, Canada, Japan, China, Brazil and the United States, into charts that show weather conditions around the world. Many of the weather maps shared on social media sites bear the Weather Bell imprint.
For years, Maue has followed the two best models in general forecasting accuracy, the European ECMWF and the US GFS model. Both models have their hits and misses with specific weather conditions, but Maue has found that the ECMWF outperforms about three percent in its five-day forecasts across a number of variables.
There are three main ways in which a forecasting model, a mathematical representation of the climate system, can improve. One way is better physics, using equations to model atmospheric dynamics that drive the movement of pressure systems around the world, generate rain and cause temperature fluctuations. Another way is to increase the resolution. A final way is to improve the recording of real-time weather data to determine the initial conditions for a prediction model run.
The new upgrade brings the biggest improvement through the second way, resolution. The model operators tripled the number of locations in their three-dimensional grid around the globe, on the surface and in the atmosphere, for which solved equations produce variables such as wind, temperature and barometric pressure. Usually this is done by throwing more computer hardware at the problem, but the ECMWF already has two supercomputer systems that are in the top 50 of the Top 500 list.
In this case, better resolution came from changing the way the three-dimensional grid is aligned on Earth, Maue said. Instead of a square box, the predicted area is divided into cubic octahedra, increasing the efficiency of the model calculations. As a result, the resolution between the grid points in the operational model, which is run twice a day to make official predictions for locations around the world, has been improved from 14 km to 9 km. This allows the model to better capture smaller weather features, such as clusters of thunderstorms.
The improvement will be even greater for the model’s ensemble forecasts. Like other global models, the ECMWF runs about 50 different times during a forecast cycle with slightly different starting conditions. This gives a range of outcomes, and by looking at all of these ensembles, forecasters can get an idea of probabilities. For example, is it likely that it will be warmer or colder than usual in 10 days? The change in the grid structure allows the ECMWF to improve the resolution of the ensemble model from 32 km to 18 km.
“These upgrades launched today not only increase the detail of our global operational forecast, reach and accuracy, but do so in a sustainable manner, paving the way for further improvement,” said Florence Rabier, Director General from ECMWF. “This was only possible because of the work that ECMWF and its partners are doing on coding and computing efficiency.”
All about the assimilation
Maue follows the improvement of the upgraded European model, which has been running in an experimental mode for several months, and found that it outperforms the operational version by about 0.5 percent. This will actually increase the performance difference by 17 percent over the GFS model.
The GFS model will receive its own upgrade in May, Maue said, in time for the start of the Atlantic hurricane season. And that upgrade could fix one of the main differences between the two best models in the world. Unlike the GFS model, the European forecast system uses a technique called 4D-Var to assimilate weather data into its forecast system, which extracts the maximum value from satellite observations and other sources of atmospheric information. Better initial conditions invariably lead to better forecasts.
The May upgrade to the GFS model will likely include this more advanced data assimilation method.