What Limits Forecast Accuracy?

Some weather forecasters make claims for accuracy, precision, availability worldwide out to 7 days ahead but we sailors know, intuitively, that the weather is never exact, so how can forecasts be precise or location specific? Of course, and to quote Professor Joad (a few of the real oldies will know the name) - it all depends on what you mean by accuracy and, precision.

Forecasting is now remarkably good and anyone who doubts that should think back to early February 2009 and the excellent forecasts of snow that affected much of Britain. Due to its maritime climate, snow is often difficult to predict over the UK. Falling at temperatures only a little lower than 0°C, the snow often has a high water content making it difficult to clear. Further, ice at these temperatures is more slippery than at lower temperatures such as occur over mainland Europe.

Particularly impressive was the last day that snow fell over southern Britain ahead of a belt of rain spreading from the southwest. Rain falling into cold dry air evaporates and cools the air bringing the melting level lower, sometimes to the surface. The rain turns to snow and can cause severe and all too obvious embarrassment to the poor forecaster.

However, on this occasion, although not unequivocal, the computer gave good advice and the human did the rest. Remember that, as I have described on another page, all weather forecasts are based upon Numerical Weather Prediction. In principle, this calculates all the physical factors driving the atmosphere.

Briefly, these are due to the spin of the earth, heating by the sun, topography and water in all its forms - liquid, solid and gas. As a onetime forecaster working prior to and during the first few years of NWP, I know full well that the human brain simply cannot even start to quantify these effects.

Had it not been for computers meteorology would have been stuck in a 1950s time warp. At that time subjective weather prediction had developed as far as it could go. To put it bluntly, but honestly, forecasts were, and could be no more than, intelligent, scientifically based guesswork.

So, as computers are continuing to develop surely weather forecasts will get better and better? Well, that is a pious hope that will simply not be realised. Let us see why there are limitations in our ability to forecast the weather.

As an approximate rule of thumb, computers double in their potential power every 18 months. In 1959 the UK Met Office had a state of the art computer capable of 3000 flops (floating arithmetic calculations per second ie additions or subtractions of numbers like 4.1468 x 10⁴ ). All calculations can be reduced to additions or subtractions By 2009, speeds were up to 16 billion flops and, by 2013 their IBM super computer will be capable of 125 trillion flops.

To help to stop your eyes from glazing over, in 1980 the European Centre for Medium Range Weather Forecasts installed a Cray 1 that was capable of some 250 million flops. About 25 years later Nokia introduced their 6300 cell phone that is capable of 237 million flops. Although it cannot fulfill all the functions of the Cray 1, it is in the same ball park in terms of raw power.

To match the atmosphere would mean increasing computer power by a factor of about 10³⁶ or 10 with 36 zeroes. That is, at best, 150 years or so away and may never happen in any case.

The Cray 1 used 115 kW of power, was cooled by liquid nitrogen and was capable of 250 million flops.

The Nokia 6300 uses 1.5 W and is capable of 237 million flops.

Lack of enough computer power and data inadequacies are serious enough but the major limiting factor to accurate weather prediction is the atmosphere itself. This results from chaos – the so called Japanese butterfly effect - the idea that a butterfly flapping its wings in Tokyo could create a storm in New York some days later.

However unlikely, that may seem, it is a fact that small weather features can grow into large systems over periods of a few days. Many of the frontal depressions that affect waters around Europe start as small waves off the eastern sea board of North America.

Precisely when and where the next wave will form is part of the chaos. As such systems develop, their precise development and tracks will depend on small variations in and near their circulations.

North Atlantic hurricane start as small groups of showers being in the right place at the right time, often near the Cape Verde islands. Very small differences in initial position and surrounding conditions can lead to markedly different behaviour.

There can be considerable differences in behaviour and considerable uncertainty, even of a mature hurricane approaching the Caribbean or the mainland of Mexico and the USA. The infamous 1987, “Mike Fish there will not be a hurricane” storm was a result of an old hurricane leaving some warm and moist air.

This came across the Atlantic before becoming embroiled in a low approaching the English Channel. There was a sudden release of vast amounts latent heat as the air was lifted and cooled. It was not a hurricane; poor old Mike Fish was quite right. Like all our lows it derived much of its energy from latent heat and, in common with a hurricane, that was the major driver.

Although one can never be sure, it is highly likely that better satellite observational data and increasingly powerful computers will mean that future storms of such magnitude will be far better, although never perfectly, predicted. One of the more impressive aspects of NWP has been the good prediction some days ahead of major storms.

The effect of chaos is to set a limit on the skill in deterministic prediction, that is the broad detail of the major highs and lows rather than monthly, seasonal or even climate prediction. It is generally thought that all skill will disappear after about 15 days. Up to that time, there can never be total certainty but skill will improve.

This is a schematic graph of skill against time in days.

For T=0, large scale analysis will be good (although that is not trues for small scale detail). By five days, skill will be appreciable but forecasts will not be completely reliable. By seven days skill will be less and forecasts rather hit and miss. By 15 days there will be no skill at all.

Global weather prediction is necessary partly because weather cam move quickly. Tomorrow's storm over the UK might not even be a twinkle on yesterday's charts. More importantly, there are long distance effects. For example, a strong NW airstream over the west of North America can create a SW flow over the western Atlantic and these, in turn can cause a NW flow over the Eastern Atlantic.

After the floods in Queensland and severe stroms in Brazil, there must be few who are unaware of the importance of El Niño and La Niña. These areas of cold or warm water result from the deep ocean circulation. They have some not yet understood (by me, at least) effects on the UK weather.

Better data and increasing computer power will allow shorter grid lengths. This will give better definition of the flow around synoptic scale weather and, therefore, better prediction of the movement and development of such large scale weather patterns.

At present, I use forecasts to six days ahead for our planning when cruising. In time to come I will expect better forecasts of large scale weather on the five or six day time scale and more useable skill to eight (at a guess) or more days.

At present, data analysis over the sea can only define weather, at best, on a scale of around 30 NM and calculation on a grid size of 1/10th of a degree ie around 6 NM. Over and near to land, topographic effects can be calculated on finer grid lengths but predictions will always be subject to constraints imposed by knowledge of small scale weather moving into an area.

In a static weather pattern and as long as the large scale pattern is well known, then small scale prediction can work well. How far ahead such predictions can usefully be made is critically dependant on the lifetimes of small scale weather patterns. Gusts of size 10s or 100s of metres in size have very short lifetimes. A small cumulus cloud, size around 1 NM will last about 30 minute. A thunderstorm, five to ten NM across will have a total lifetime of a few hours.

Groups of storms may last a day or two. These figures give some idea of small scale predictability and explain why the Met Office only calculates detail on a 6 NM grid to 48 hours and only uses the output to some 30 hours. That is despite having the best data analyses currently possible; few organisations making short term, detailed predictions start with as good an analysis as that used by the Met Office.

In time, Met Office detailed forecasts will allow better prediction of small scale topographic effects, such as localised strong winds near headlands and through narrows such as the Dover Strait. They will give better prediction of squall lines and of severe, small scale weather such as super cells. Time scales are likely to be one or two day ahead.

One of the benefits of increasing power will be the use of ensemble techniques. The numerical weather prediction model can be used to see where data sensitive areas are located. That is areas where small changes or errors in the data can give a very different forecast. By putting in random variations in the data, the model can be run several times and a spread of results obtained. This can then be used to derive probabilities of different forecasts.

A good example was in the snow of February 2009 over the UK. On the last day of snow in the south, the ensemble suggested a 20% probability of snow. What would London Councils do with that information? What would the press do? Banner headlines - SNOW IN LONDON say Met Office! Followed, no doubt, on the day wfter by bad reviews.

In the event, the forecasters decided to play the odds and predict no snow in London. A wise and well justified decision but it does highlight the question of who should be given probabilities of this kind.

Just how a sailor would use probabilities is uncertain and one reason why the Met Office has never gone down that path in it published forecasts despite some public pressure. The forecasters can use the information to express doubts and uncertainties, I suspect that this is just as useful to a sailor who would not really know what to do if told that the chance of a gale was, say, 30%. The ensemble technique can be used both for large scale and local detailed predictions.

However good forecasting is, there will always be the problem of communication with the user. To get a feel for this, anyone who criticise forecast texts should do a little experiment. Sail along a coast for 12 hours, say 6 one way and 6 back. Note all the weather and winds. Then try to describe them in a text that could be heard or read easily and understandably. Then imagine doing that for a length of coast from Lyme Regis to Lands End for 24 hours with all the uncertainties of a forecast and bearing in mind all the local coastal effects, headlands and such like.

Do not believe such claims as “…real-time, accurate forecast data”. “...animates the weather very accurately and locally from 1 hour up to 7 days out for any location on the globe” or “...weather forecasts of exceptional detail and quality”. Take all such with a big pinch of salt. Remember that little is absolutely certain in any weather forecast, except for the date. Even then you must keep the theologians apart.

Weather forecasting will improve but will never be perfect.