Slashdot Mirror
Better 'Nowcasting' Can Reveal What Weather is About To Hit Within 500 Meters (technologyreview.com)
Weather forecasting is impressively accurate given how changeable and chaotic Earth's climate can be. It's not unusual to get 10-day forecasts with a reasonable level of accuracy. But there is still much to be done. One challenge for meteorologists is to improve their "nowcasting," the ability to forecast weather in the next six hours or so at a spatial resolution of a square kilometer or less.
From a report: In areas where the weather can change rapidly, that is difficult. And there is much at stake. Agricultural activity is increasingly dependent on nowcasting, and the safety of many sporting events depends on it too. Then there is the risk that sudden rainfall could lead to flash flooding, a growing problem in many areas because of climate change and urbanization. That has implications for infrastructure, such as sewage management, and for safety, since this kind of flooding can kill. So meteorologists would dearly love to have a better way to make their nowcasts. Enter Blandine Bianchi from EPFL in Lausanne, Switzerland, and a few colleagues, who have developed a method for combining meteorological data from several sources to produce nowcasts with improved accuracy.
Their work has the potential to change the utility of this kind of forecasting for everyone from farmers and gardeners to emergency services and sewage engineers. Current forecasting is limited by the data and the scale on which it is gathered and processed. For example, satellite data has a spatial resolution of 50 to 100 km and allows the tracking and forecasting of large cloud cells over a time scale of six to nine hours. By contrast, radar data is updated every five minutes, with a spatial resolution of about a kilometer, and leads to predictions on the time scale of one to three hours. Another source of data is the microwave links used by telecommunications companies, which are degraded by rainfall.
Their work has the potential to change the utility of this kind of forecasting for everyone from farmers and gardeners to emergency services and sewage engineers. Current forecasting is limited by the data and the scale on which it is gathered and processed. For example, satellite data has a spatial resolution of 50 to 100 km and allows the tracking and forecasting of large cloud cells over a time scale of six to nine hours. By contrast, radar data is updated every five minutes, with a spatial resolution of about a kilometer, and leads to predictions on the time scale of one to three hours. Another source of data is the microwave links used by telecommunications companies, which are degraded by rainfall.
"It's not unusual to get 10-day forecasts with a reasonable level of accuracy"
A UK 10-day forecast consists of the words "The sun is likely to come up; you may or may not be able to see it."
5-day forecasts are generally little better than flipping a coin to see whether it will rain or not. 3-day isn't too bad and 1-day forecasts are reasonably good for much of the summer and winter; in spring and autumn they're pretty rough.
None of which prevents the Met Office from showing weather maps with a ludicrous level of precision completely unmatched by their accuracy.
"Encyclopedia" is to "Wikipedia" what "Library" is to "Some people at a bus stop"
This story is nonsense. I'm a meteorologist and I do severe storms research. There are a number of factual errors present, even in the summary. Nowcasting is a short term forecast, generally in the 0-6 hour time frame. That's one of the few things this story got right.
Forecasters rely heavily on numerical models to make predictions. On a regional scale, these models numerically integrate a number of partial differential equations forward on a 3D grid. Many of these models are different configurations of the Weather Research & Forecasting (WRF) model. WRF can be run across many cores with shared memory (OpenMP) or distributed memory (MPI). Domains with very large numbers of grid points can be run across thousands of cores. If the spatial size of the domain size remains the same, adding grid points means decreasing the space between each grid point. Not only does this increase the processing requirements because of more grid points, but also the time step of the numerical integration generally has to decrease. High resolution domains require very large amounts of computing resources in order to produce a forecast in a reasonable amount of time.
The highest resolution model that's regularly run operationally in the US is the High Resolution Rapid Refresh (HRRR) model, and is a specific configuration of WRF. The HRRR is run hourly and has a horizontal grid spacing of 3 km. This is well above the supposed precision of 500 meters. Furthermore, even if the HRRR was run at 500 m, it doesn't mean the forecast would be accurate on such small spatial scales. The big difference between the 3 km HRRR and coarser resolution models like the 13 km RAP (also, WRF-based) is that the HRRR doesn't parameterize convection. That means it runs at a high enough resolution that it can directly simulate phenomena like thunderstorms.
The resolution of radar data in the US is about 500 m, and has been for roughly the past decade. The best weather satellite right now is GOES-16, with a resolution of 500 m-2 km, depending on the type of product. That's a huge difference from what's described in the summary. Forecasts that rely on extrapolating radar and satellite data might be accurate for 30 minutes or perhaps even an hour or two. Beyond that, numerical models are going to produce better forecasts.
The radar and satellite data, along with a lot of other data sources, are assimilated into models like the RAP and HRRR. Assimilation basically means updating the state of the 3D domain based on the new observations. Data assimilation of conventional observations like winds, temperature, pressure, and humidity generally produces good results. However, assimilating radar and satellite data isn't as simple.
Reflectivity and radial velocity are generally assimilated from radar data. Radial velocity is generally assimilated in areas where there isn't precipitation, and is a lot like assimilating wind data. Reflectivity is the amount of power that's scattered back to the radar, and is generally larger if there's heavier precipitation. It's not nearly so simple to assimilate reflectivity because you also need to update variables like temperature, wind, humidity, and pressure in the 3D domain, even though the radar isn't directly measuring them. Those variables are going to be quite a bit different inside a thunderstorm than they are outside it. If you want to update the position and strength of thunderstorms in a model, you need to update quite a few variables in the model that you probably aren't measuring at all in those areas. If you want accurate forecasts of thunderstorms, you need to update the model based on radar reflectivity data.
There are techniques like the Ensemble Kalman Filter (EnKF) that update unobserved variables based on measurements of variables that are observed. However, even with the best EnKF techniques at present, assimilating reflectivity data often doesn't really improve the forecast beyond an hour or two. Perhaps more observations and better techniques will im