Medium range weather forecasting

Pitfalls and solutions

By Brian Gaze

Medium range forecasting pitfalls

Medium range NWP deterministic computer models produce outlooks for up to around 16 days ahead. They are widely used to generate forecasts on websites, including TheWeatherOutlook, and apps. However, when looking more than a few days ahead the accuracy of individual model runs quickly starts to fall.  

"Numerical weather prediction (NWP) uses mathematical models of the atmosphere and oceans to predict the weather based on current weather conditions"

Will the weather forecast be right?

When looking more than one week ahead there can be massive differences from run to run. To illustrate the point here are two charts generated using data from the Global Forecast System (GFS) model. The GFS is a medium range numerical weather prediction model run by NCEP in the United States. It is considered to be one of the best available, along with others such as those operated by the European Centre for Medium-Range Weather Forecasts (ECMWF) and the UK Met Office.   

The first chart is for 384 hours ahead from the initiation time. It shows a plunge of very cold arctic air moving southwards over the UK. If it was correct there would be a risk of snow in central and northern counties and possibly even in parts of the south. Maximum temperatures would range from around 2C to 8C.

The second chart is for 372 hours ahead and it was generated from the next but one run of the GFS model. Therefore, it is valid at the same time. This time a deep area of low pressure is centred in the Atlantic and a warm sector associated with it is crossing the UK. 

Instead of a cold scenario with the possibility of snow, this shows a very mild and wet picture. Temperatures range from 6C over the Scottish mountains to 16C near to the south coast.

Therefore, within the space of a couple of runs things have been turned on their head. This is quite an extreme example and at shorter ranges, for example one week ahead, there is less variability from run to run. Nonetheless, it can still be very significant and have a big impact on the forecast which appears on the website or app you are using.

So can anything to be done to improve medium range forecast accuracy with current technology? In short, the answer is yes it can by using what are known as ensemble models.

What are ensemble models?

An ensemble is also a Numerical Weather Prediction (NWP) model, but instead of being run once it is run many times with the starting conditions varied or perturbed slightly each time. The reason for doing this is because it is impossible to feed enough data into the computer models to tell it the exact state of the Earth's oceans and atmosphere at the start time. Therefore, making slight alterations helps to account for the uncertainty that is present.

How to use ensemble models?

A range of ensemble models exists. Examples are the NCEP Global Ensemble Forecast System (GEFS), the UK Met Office MOGREPS-G and the ECMWF Ensemble. 

Data from ensemble models is plotted in a number of different ways. Below are some typical examples. 

Types of ensemble model charts

Firstly, mean chart can be generated for a given time. These are calculated by averaging out a forecast variable, e.g. pressure, from all of the runs in the model. One advantage of doing this is that extremes or outliers are smoothed out and a blended picture is produced. A disadvantage is it can mask possible scenarios and even suggest unsupported ones.     

Sometimes instead of looking at the average it makes more sense to view all of the individual runs to identity clusters. There are also automated ways of doing this. However, a manual approach may involve viewing the "postage stamp" charts, such as the one below.

Each stamp shows the forecast from one run in the ensemble at a given time. The chart below is showing pressure patterns, but other examples would include temperatures, wind speed and rain rates.

Another frequently used approach is to plot line graphs showing forecasts for a particular location over the time period. For example, the 16 day GEFS plot below is for London. Time is on the horizontal axis with air mass temperatures and rainfall on the vertical ones. 

A key advantage of this type of plot is it provides a quick way of seeing what all of the individual runs are favouring for the given location. Also, it is common to plot the ensemble mean to provide additional guidance. Other data such as the 30 year mean and the deterministic model forecast, e.g. the GFS, can be added to provide more insights. 

On the graph it is possible to see that the GFS (shown by the thick green line) was forecasting warmer air mass temperatures than the ensemble mean and most of the individual runs during the second half of the 16 day period. In addition, the ensemble mean, shown by the thick purple line, is well above the 30 year norm, shown by the thick black line. 

Making a more accurate forecast

The key things to consider are that the further ahead the forecast is for the less accurate a single deterministic model run will usually be. Therefore, the ensembles come into their own to help identify the likelihood of different outcomes.

However, they are not a silver bullet and although they are much less prone to massive swings than the deterministic runs they can change quickly. Most importantly, when looking more than about seven days ahead it is about identifying trends and the probabilities of different outcomes.

You can make use of ensemble data directly by using the model charts discussed above or data tables - GEFS data tables, MOGREPS-G data tables, ECMWF ENS data tables.

The Will it snow, Will it rain and Will it be frosty forecasts make use of ensemble data to increase accuracy at longer ranges. 

News and information

RECENT WEATHER NEWS