Impact Case: Small Region, High Resolution, Local Weather Forecasts

Summary of the impact

 Two prototype systems by the Department of Computer Science, UiT, have impacted weather forecasting in Norway in three ways.

The first impact is due to a prototype wall sized high-resolution tiled display and visualization system which helped meteorologists detect errors in existing weather models. Removing the errors has impact on the accuracy of public weather forecasts.

The second impact is due to a protype “personal” weather forecasting system which does very high-resolution forecasts for very small areas in just a few minutes on a typical personal computer. The grid size is reduced from the typical 2.5-8 km to a few hundred meters. Doing many such forecasts is not practical timewise even on a supercomputer. However, doing it by enabling anybody to do a forecast on their personal computer scales well with the number of forecasts needed. Further, “stamp-sized” high-resolution forecasts result in more accurate forecasts wherever there are prominent landscape structures, like around mountains and in fjords. This has potential impact in several use domains, like planning for selling electricity from windmills, and for doing weather forecasts for airports.

The third impact is due to the other two. The present director for all weather forecasting in Norway, Dr. Bård Fjukstad, was directly involved with the research behind both prototypes, and it has impacted how he approaches present and future weather forecasting which will use stamp-sized forecasts to provide the "ground truth" for AI models to train on.

Underpinning research

By using a prototype wall-sized high-resolution display wall and an associated visualization system, errors in two existing weather forecast models were discovered by meteorologists at the Division for Forecasting Tromsø (“Værvarslinga for Nord Norge”), Forecasting Department, Norwegian Meteorological Institute. These errors were reported and subsequently removed from the forecasting models.

The prototype display wall comprised 28 PCs, each with a projector for a total of 22Mpixels. The projectors were placed in a 4 by 7 grid for a diagonal size of 220 inches. The system included a compute cluster of 30 computers to store and process content. The user interface comprised 17 cameras and four microphones to track visual and aural multi-user gestures for controlling applications. User input was touch free and could be given both close to the display wall, and from a distance. To let the users of the display wall display output from whatever application they wanted, but at larger size and with more details, a novel way of using the virtual network computer (VNC) protocol in a tiled way was invented.

Multiple weather models are run every day to produce input for a public forecast. The grid cell walls are typically quite large at 2.5 to 8 km. Because of this, the impact of ground structures, like mountains, on wind magnitude and direction, are not accurately reflected in the forecasts. Computing forecasts for smaller cells are not scalable in that it would take too long time to do so for all of Norway with the available supercomputers. Instead, to make it scalable to produce an accurate forecast for a small cell anywhere in Norway, whoever needing such a forecast can do it themselves on-demand at any time and within a few minutes. To enable users to do this, the prototype comprises an implementation of a weather model suitable for a typical PC, the location coordinates of the relevant small area, and the background large cell forecast provided by the meteorological office.

The prototypes were researched and developed by the High-Performance Distributed Systems (HPDS) group (later re-named to the Cyber-Physical and IoT Systems group, CPS) at the Department of Computer Science (IFI), University of Tromsø (UiT).

Details of the impact

Prototyping a system is done to be able to document its behaviour through scientific experiments measuring relevant performance metrics. This defines a basis for comparing the system with other systems and makes it possible to document if and where the prototype does better. When a prototype is applied in some use domain, like weather forecasting, it becomes a new instrument for the domain and can result in unexpected insights.

Today, meteorologists explore a forecast on typically three displays placed side by side. The problem with smaller displays is that when zooming out to larger areas, like Norway, the finer details of the forecast are not visible, and when zooming in, the larger structures are not visible. However, on the prototype display wall, the large size and the high resolution enabled multiple meteorologists to simultaneously explore the forecasts by physically moving away from the display wall to see larger structures and moving closer to explore the finer details. This was made even more efficient by a user interface allowing simultaneous input from multiple users and from anywhere in the room.

The errors in two of the existing weather models were discovered when using the display wall because the meteorologists could view both the overall structure of the forecast, and the finer details, at the same time. This made it simpler to see artificial repeating errors in the forecasts masked by smaller displays.

Doing weather forecasts for Norway with grid cells significantly smaller than 2.5 km is presently not practical because the forecast is not ready fast enough using the available supercomputer(s) for a public forecast to be made in time. However, forecasts with grid cell wall sizes of, say, 500 meters, are not needed everywhere. Instead of computing such forecasts for Norway, they can be computed where they are needed. While it is known a few places where they are needed, like around airports, in most cases, this is not known ahead of time. Instead, computing forecasts can be done where they are needed and by those who need them. The prototype system did this by combining a tablet with GPS sensors and a PC. The user can click on a map on the tablet for where the forecast is needed, and the tablet part of the prototype forwards this to the PC where the corresponding forecast is computed in a few minutes. The result is sent back to the tablet and the forecast is displayed onto the tablet (the visualization of the forecast can even be done in an augmented reality way).

The impact of the small area forecasting system is both in off-loading the computing needed to do the forecasts from the central resources of a met office to many personal edge computers, and in providing for both on-demand and more accurate forecasts.

More accurate forecasts reflecting land structures have value in many domains, including in planning windmill energy production and selling, for autonomous electric ships and ferries, for sailing in fjords, and for rescue operations in mountain areas.

Sources to corroborate the impact

The director of the Forecasting Department, The Norwegian Meteorological Institute, was involved in the research and can be contacted to verify the claims made with regards to the impact it has had on weather forecasting:

Dr. Bård Fjukstad
Director of Forecasting Department,

The Norwegian Meteorological Institute.
Postboks 6314 Langnes, 9293 Tromsø, NORWAY

Email: b.fjukstad@met.no

Telephone: +47 48 23 46 83

References to the research

Fjukstad, B., Hagen, TM.S., Stødle, D., Ha, P.H., Bjørndalen, J.M., Anshus, O. (2012). Interactive Weather Simulation and Visualization on a Display Wall with Many-Core Compute Nodes. In: Jónasson, K. (eds) Applied Parallel and Scientific Computing. PARA 2010. Lecture Notes in Computer Science, vol 7133. Springer, Berlin, Heidelberg, 2012,            https://doi.org/10.1007/978-3-642-28151-8_14

Bård Fjukstad, John Bjørndalen, Otto Anshus, Accurate Weather Forecasting Through Locality Based Collaborative Computing, The 8th International Workshop on Trusted Collaboration (TrustCol 2013), Austin, Texas, USA, 2013, https://eudl.eu/doi/10.4108/icst.collaboratecom.2013.254178