Weather model – a modern forecast factory

8.9.2016 12:52

Weather model is like any computer programme that consists of a complicated code for the computer to process in order to generate the desired outcome. However, this simple explanation does not do justice to or value the research and work that the development of weather models to their current level has required.

Photo: Antonin Halas

The history of numerical weather models can be considered to have started in 1922 when Lewis Fry Richardson, an English mathematician, tried to forecast weather numerically for the first time. "The basic principles that he developed have remained the same until the present day," says Sami Niemelä, Head of the Meteorological Research Unit at the Finnish Meteorological Institute (FMI).

A weather model is a description of the entire system of the atmosphere. It is based on the physical principles of conservation (mass, momentum, energy and the different phases of water) on the basis of which a group of equations is formed and an attempt is made to solve it numerically. The forecast is launched by creating the initial state of the atmosphere by using weather observations along with other a-priori information. Richardson already divided the globe into a so called grid that consists of thousands and thousands of calculation points in the atmosphere both horizontally and vertically. The equations reveal how, for example, the temperature or wind changes as time progresses in each gridbox. The result obtained will always be used as the starting point when the next calculation step is started, until the end of the required forecast period has been reached.

Richardson tried to calculate a 6-hour weather forecast for two points by solving equations with help of just a pencil and paper. It took Richardson six weeks to calculate all those equations, and the final forecast was not very successful, either. Richardson realised very soon how impossible the task was, but he already had a vision of "forecast factories" of 64,000 mathematicians solving equations together in order to forecast the weather across the world.

Development of computers enabled introduction of weather models

Richardson's visions were not realised until decades later when the development of computers made it possible to create the first simple weather models. With help of computers, it became possible to perform millions and millions of calculations fast and accurately enough. It is only with help of efficient super computers, i.e. the modern day "forecast factories", that it is possible to calculate the changes in different weather variables such as temperature, humidity or wind speed in the following hours, days, weeks and today even months at different locations of the forecast area from sea level to the height of dozens of kilometres. "Weather forecast models are therefore not only statistical models that are based on an observed state only like the time series models used in economics," Niemelä stresses.

Today's weather models have been divided into both global and regional models that serve different purposes. Finland utilises the global model of the European Centre for Medium-Range Weather Forecasts (ECMWF), which is the best model in the world in its own category. A global model describes the general circulation of the entire atmosphere and creates a broader outline for weather forecasts at global level.
The distances between the grid points in the European Centre's model are currently about 9 kilometres. In order to obtain more accurate information from a specific area, such as Finland, separate limited area models are needed. Although the resolution in regional models is better, they cannot work without the boundary conditions provided by a global model. On the other hand, a model for a smaller area can be run several times a day while taking advantage of the most recent observations. Furthermore, such models can be modified more easily to meet the special requirements of the target areas.

Developing a model requires collaboration

Development of models requires a huge amount of computational resources as well as diverse expertise. Therefore, different countries have formed and joined different consortia in which weather models are developed in cooperation. Finland and the other Nordic countries have already been members of the Hirlam consortium for 30 years. At the beginning of 2004, the Hirlam consortium started close collaboration with the Aladin model consortium originated from Meteo France.  The HARMONIE model, which is capable of resolutions of up to kilometre scales, was created as a result of this collaboration. It is currently in operational use at the FMI. Finland's contribution to the model has been expertise related to both data assimilation methods and parameterization of physical processes; especially to how the interaction between the Earth's surface and the atmosphere works, how radiation, turbulence in the boundary layer and cloud microphysics affect the state of the atmosphere and how this information can be efficiently used in the model.
"The weather conditions that prevail in Finland are challenging, but the model must be able to take local conditions into account. HARMONIE is capable of describing, for example, snow bands that are formed in the Gulf of Finland because the model can take into account the impact the unfrozen sea and, for example, Lake Ladoga have on weather," Niemelä explains.

Different models have different development objectives

The accuracy as well as the length of weather forecasts have developed continuously. Global models attempt to stretch predictability even further. The development has been stable and forecasts have improved by one day  in a decade on average, so a five-day forecast today is effectively as good as a four-day forecast was 10 years ago. There is now a tendency to stretch the limits of predictability into a second week or, like in seasonal forecasts, even into months. "There is a huge societal demand for this kind of forecasts at the moment; that is why so much is invested in developing them in terms of global models," Sami Niemelä says.

"As regards limited area models, the development has a different scope. These models focuses on short range forecasts (1–3 days) or very short range forecasts (<1 day), which are based on assimilation of high resolution observations and which aim to carry on the signals obtained from the observed information yet more accurately into the following hours. This information is important, for example, for the smooth running of air traffic."

Predictions for forecast uncertainties

The traditional idea in weather models is that the model will give one answer, i.e. what the weather will be like in a specific location at a specific time. "However, it is a fact that models will always be simplifications of the real atmosphere." Niemelä reminds us that, although the current super computers make quintillions of calculations per second, no model can ever perfectly copy the full complexity of reality.

In his opinion, there are two reasons for why we cannot obtain perfect weather forecasts. "On the one hand, the starting point that weather models need is the most accurate description of the atmospheric state." However, it is not possible to obtain an exact description as weather observations are not available everywhere and, occasionally, there may also be errors. Even small errors in the initial conditions may cause large errors in the final forecasts. On the other hand, there are also uncertainties in the modelling methods. "That is why it is not sensible to give one answer, but in addition the models should be able to give estimations about the uncertainty of the predictions – in other words, to predict the forecast uncertainty."

The so called probabilistic forecast systems are indeed a current trend in weather models. In these systems, the initial state and the evolution of the actual forecast are perturbed in order to calculate different possible alternatives, i.e. so called ensemble predictions are made. They can be used to draw up a precise weather forecast, and also to estimate the weather dependent uncertainty of the information produced by the model. "The weather model can produce a very fragmented group of predictions in certain weather situations, in which case the weather forecast is expected to be very uncertain. In other situations, the ensemble of predictions is very uniform and the meteorologists can to a large extent rely on the accuracy of the weather forecast, which is of course our main objective, " Niemelä summarises.