Numerical Weather Prediction (Oct 2017)

Just as you were thinking that you had had enough of three-letter acronyms, here is another one: ‘NWP’ or Numerical Weather Prediction in full. In recent years there have been some great advances in NWP that are of real benefit to the private IFR pilot looking to reduce planning stress and improve transport utility of light piston GA. However, like other technological developments, NWP is moving faster than the GA training industry can change.

Having an instrument rating or an IR(R) means being able to deal with a much wider range of weather than the VFR pilot. However, I would argue that the instrument pilot flying light piston GA needs an even deeper understanding of weather phenomena and of weather forecasting than the VFR pilot. The main reason for this is that we often expect to use our qualifications and our aircraft to achieve something as close as possible to the utility promise of commercial air transport: going where you want, when you want. The “when you want” part of this promise is all about the weather. If you are willing to wait long enough, you can fly any route in the world in VMC. However, we have invested a lot of time and money in order not to have to wait. We also plan trips, business or pleasure, where we need to know in advance with some confidence that we will be able to make it there safely. If it looks like this is not possible, it may be time to start booking scheduled tickets or cancelling hotels. We may also want to know in advance if we are likely to be able to get back and maybe do so in a way that will provide a comfortable trip for some non-pilot passengers.

Perhaps that is the main difference between weather forecasting for VFR pilots and IFR pilots. The IFR pilot is able to deal with a much wider range of weather but may well need to know, with a degree of certainty in advance, if a mission is flyable. The secondary difference is that the IFR pilot will be choosing to penetrate or avoid certain weather where understanding its spread and intensity is critical to flight safety. For these purposes, NWP is your friend. There are some amazing resources available that greatly aid flight planning in advance and in setting flight strategy on the day and most of these resources are free. However, because it is relatively new, the use of NWP is not covered in theoretical knowledge exams and because of the speed of change it is unlikely that TK will catch up. My personal view is that topics such as this are correctly covered by Continuing Professional Development (CPD) seminars such as those offered by PPL/IR Europe on topics such as Performance Based Navigation.

Like so many other things in aviation, it is also important that the pilot has some grasp of the fundamentals and of the limitations of NWP in order to be able to use the tools safely. I think of this as being like the need to know about coastal effect or operating range before using an ADF to fly an approach. The objective of this article is to provide a grounding in the principles and the sources of NWP-based weather products. Its aim is to increase awareness of what NWP can and cannot do. It will not cover forecasting of specific weather phenomena or strategies for flying them. For that you will need to come to the next PPL/IR Europe weather seminar!

By way of personal introduction, I did my instrument rating back in 1993 and had the good fortune to be working for an amazingly enlightened consulting firm who thought it perfectly fine for me to be flying myself to business meetings and sending them the bill. This was long before the low cost airlines, or online booking for that matter. I needed to be able to be sure I could make the trip. At the same time, I was sharing an aircraft with someone who went on to found one of the first online aviation weather businesses. All this combined to me developing a bit of an “unhealthy level of interest” in weather, even for a pilot.

Why is weather forecasting so hard to get right? Pilots and weather forecasters have had a strained relationship since the very early days of aviation. The reality is that weather forecasting, and especially for pilots of light aircraft, is a difficult thing to do with accuracy and with certainty. To understand why this is so, we need to start with a fundamental appreciation of what weather is. Our planet is covered with a layer of fluid that we call air. Air at the poles is cold and tends to sink down towards the equator. Air at the equator is warm and tends to rise towards the poles. How these masses of air mix together and also mix with moisture define weather.

The first issue is all about what happens as those masses of air mix. They do so in a very messy way. The technical term is that the mixing is turbulent fluid flow. One example of turbulent flow is a fast flowing stream, with the water twisting and turning and crashing around stones. The opposite of turbulent flow is laminar flow, when the fluid flows in a nice orderly smooth way. Continuing to use a water analogy, think about a large, slow moving river. Turbulent flow is chaotic and seemingly random. Imagine how difficult it is to predict in advance exactly what chaotic pattern will happen next.

The second issue is that, as pilots, we need not just a prediction of what will happen next in this chaotic pattern but we also need that prediction to be made with a very high degree of accuracy. A change of just 2 degrees centigrade or of one hour in the timing of the arrival of some cold air can make the difference between conditions that are perfectly flyable and those that are dangerous. There are several mathematical equations of motion that set out to describe what is going on with turbulent fluid flow and many of these equations have been around for decades. The most notorious are the Navier-Stokes equations, which date from the 19th century but are yet to be fully solved. They are listed as one of the seven most important outstanding mathematical problems. This is the core issue with weather forecasting. If only we could apply equations of motion to the masses of air encircling the earth, we could describe mathematically what will happen next and do so with definitive precision. As we are not able to do this, we need an alternative. This is where modern computers come in with NWP. Instead of trying to solve the impossible, NWP uses the number-crunching capabilities of a computer to simulate the world’s weather.

Setting up an NWP model

An NWP model starts with the development of a grid. The entire atmospheric layer surrounding our planet is mapped out in three dimensions. This is just like the latitude and longitude mapping with which we are familiar for navigation, but with the additional dimension of altitude, or to be precise, pressure altitude. Instead of a series of roughly rectangular-shaped grids on a chart, we instead have covered the planet with a series of roughly cuboidal blocks. These blocks are called “gridcells”.

It is worth considering for a moment some numbers relating to gridcells. These numbers get quite big, and this has an impact on NWP but as we will see later, these numbers would benefit from being even bigger.

A typical gridcell might be 0.25º of latitude and an average of about 0.25º of longitude in area. That makes about 2 million blocks. It is then typical to model the atmosphere vertically in 64 layers going all the way up to 1 mbar or about 100,000 feet. This yields a total of 133 million gridcells. The second step is to develop a computer model of the atmosphere based upon theories of how weather behaves. The coding of the model is such that it will use current and past actual observations to simulate today’s weather and extrapolate the model forward to create forecasts. This is a very complex task and of a huge scale. Not only are there 133 million gridcells to be modelled past, present and future but what happens in one gridcell will have a knock-on effect to the next. To give a sense of the scale of the computing task, I am typing this article on my nice new “quad core” iMac. The main computer at the UK Met Office has a quarter of a million cores and needs about 4MW to power it. It is the 11th most powerful computer on earth and many of the other very powerful computers in the world are also used for weather modelling.

The point about the knock-on effect and extrapolation is an important one for pilots to appreciate. Given the requirements of our missions, we would like a forecast to be as accurate as possible, as far out as possible. To figure out what the weather is going to do in the next ten minutes does not require a supercomputer. However, with millions of gridcells all interrelated to each other, tiny errors in observations or in the model will mean that any forecast will at some distance into the future become too inaccurate to be of use. Moving this point forward in time is one of the most important drivers in the whole met industry. I will cover this point in more detail later in this article. The third and final step is to incorporate a learning loop so that the model can improve its performance over time.

Providers of NWP models

There are a dozen or so of providers of NWP models in the world. Examples are ECMWF (European Centre for Medium-Range Weather Forecasts), JMA (Japan Meteorological Agency) and the UKMO. I think of them as sitting somewhere on the boundary of a state-funded academic institution and a commercial enterprise. Luckily for us, they all collaborate with each other in the goal of producing better forecasts, but there is also a real spirit of competition driven by professional pride. An example of such competition was the predicted path of hurricane Sandy in 2012. Both the ECMWF and UKMO predicted accurately that the superstorm would hit New York rather than dissipate in the north Atlantic as forecast in the US by their GFS model. This error was believed to be due to better equipment being available in Europe and led to a boost in funding in the US. This is good news for us. As you would expect, setting up and running an NWP model is very expensive, with budgets typically in the hundreds of millions of pounds per year.

Possibly the most important NWP model for a pilot is GFS (Global Forecast System) from the National Weather Service in the US. GFS is a work of the US Government, and under US law, is available for free in the public domain. There are parallels with GPS in this respect. As a result, many of the forecasts we use as pilots are based upon GFS, where a specialist provider takes the raw GFS data and turns it into something useful to us.

Operating an NWP model

Of most importance and interest to a pilot is how an NWP model is operated each day. There are three steps: initialisation, computation, and presentation. Because of the attractiveness of GFS being in the public domain, I will mainly use GFS as an example to illustrate the process.

Initialisation is where the model is populated with data from real observations. These observations include the obvious, such as air temperature aloft at different levels, and surface pressure. They also include plenty of less obvious parameters such as sea ice, ocean currents, snow cover and even soil moisture content. All such parameters have an effect upon how future weather will develop. Observations are made from a variety of sources such as ground stations and radiosondes. It is worth noting that the number of observations is far smaller than all the parameters in all the gridcells. Computation is where whatever observations are available are put into the model and then the model is run. As well as computing the model output, there is also a process of parameterisation. Although there are millions of gridcells, they are still quite a bit bigger than some weather phenomena and parameterisation helps with, amongst other things, low level and cumuliform cloud forecasting.

The GFS model is run at 00Z, 06Z, 12Z, and 18Z each day and the output from GFS starts to come out a few hours later. Why a few hours? Despite the model running on one of the world’s most powerful supercomputers, it just takes this long to crunch all the data. The output is called a “dataset” and it is this dataset that is freely available in the public domain. Once the run is complete, the GFS dataset provides a forecast for T+3 hours, which is normally in the past by the time it is released, out to T+384 hours, or 16 days out. The accuracy of the forecast is not the same at T+3 hours as it is at T+384 hours! More about this later.

Presentation is the final step. The dataset may well be the result of much human and computational endeavour, but it is not of any use to a pilot looking to plan a flight. Further computation is required to interpret the dataset and present information to help the pilot, such as cloud layers, precipitation or icing. There are a huge number of resources of valuable met information that use the GFS dataset. Many are of general interest and some are specifically for aviation. Reflecting the freely available dataset, most, if not all, such resources are free of charge to the end user. When I deliver the PPL/IR Europe weather seminar, I find this attribute tends to be much appreciated by pilots.

Examples of weather information based on GFS

Given how many sources of weather information are out there, the real task in hand is to choose a format that suits you, rather than there being a “best” source. After all, they are often just different presentations of the same weather model. It is perhaps worth classifying the available material into two groups: horizontal charts, and vertical cross sections. For horizontal charts, I like very much the “Expert Maps” available at http://www.weatheronline.co.uk; I also like the charts available at “Top Karten” available at http://www.wetterzentrale.de. Some of this site is in German but you only need to know a few words. To be honest, I prefer the original Wetterzentrale site, which is still maintained at http://old.wetterzentrale.de.

At these sites you can find charts for pretty much anything you may be looking for as a pilot. Cloud cover at various levels, likewise winds, CAPE and lifted index charts, precipitation and convective precipitation, freezing level and so on. Two favourites of mine are both from the old Wetterzentrale site. I have always liked the “3h niederschlag” chart, showing precipitation, convective precipitation and freezing level all on one chart. I like also the way they depict CAPE and lifted index together. However, these are my favourites, you will find what really resonates with you. You will note that these sites also offer charts from other weather models, such as ECMWF, which are not freely in the public domain. However, free charts from these models are of a much smaller range of parameters than GFS. You can get a few more by subscribing.

A vertical cross section is normally a depiction of clouds and weather along a defined route with a specified start and finish time. The first such depiction came from Ogimet http://www.ogimet.com who, I believe, coined the term “gramet aero”. The Ogimet interface is quite fiddly, particularly when setting up timing. In my view, a much superior gramet is available from Autorouter, developed by PPL/IR Europe members Achim Hasenmuller and Thomas Sailer. The Autorouter gramet can be driven independently or from the Autorouter route finding engine.

How accurate are the forecasts?

The primary factor influencing accuracy is the range of the forecast. Every presentation based on NWP should be datestamped with the time the model was run. The range is the difference between the time of the forecast and the time the model was run. Range is sometimes expressed as T+24, T+27 and so on. As you would expect, accuracy falls off with increasing range. The predominant objective today for the entire meteorological industry is to increase the forecast range for a given level of accuracy. Accuracy is constantly measured, parameter by parameter, by comparing forecast with actuals. Collectively, the met industry is increasing range by about one extra day for each decade of development. This means the T+120 forecast today is of similar accuracy to the T+96 forecast of 2007.

For the parameters that concern us for light piston IFR flying, I personally take the view for my own flying that T+72 is perfectly accurate enough for planning purposes. If I am doing a “got to get there” flight, and the T+72 forecast is OK, I will not worry about making contingency plans. Of less direct relevance to the private pilot, but to include some real data, a benchmark used by the industry is the forecast accuracy of the height of the 500mb level. By 2013, this was being forecast at a range of three days at an accuracy of 98.5%. Whenever using an NWP product, the pilot must be completely clear about the initialisation time and the range of the forecast.

What other issues relating to NWP that a pilot should understand?

One particular area of weakness is the forecasting of low level cloud near the ground, known as “boundary layer” cloud. This is mainly due to the resolution of the computer model of the earth’s surface. In contrast, the forecast accuracy of, for example, the 700mb wind and the freezing level is very good indeed. As a result, NWP is much more useful as a tool for IFR planning than for VFR planning. There are weather products offering information on lowest cloud base but these should be treated as general guidance only. Forecasting icing is still relatively inaccurate and with a significant tendency towards the false positive. The reason for this is that icing is capricious and difficult to predict but the associated NWP algorithms are still quite crude. Work is going on in the US to develop better algorithms but this will be a slow process. Few people are involved and scarce PIREPS are needed to provide the data to prove or disprove accuracy.

Paradoxically one issue is the quality of the presentation graphics. This is getting better all the time and is a great visual aid for the pilot. However, the picture looks just as pretty at T+384 as it does at T+24, despite the T+384 forecast of being of no aviation value at all. There is also a tendency for some forecasts to imply more cloud detail than is actually being produced by the model. Finally, most of these weather products are fast moving and nearly always free. There is no formal service level agreement behind them as there would be for an official aviation met provider. I find the quality of service provision to be excellent but everyone should have a backup source should their preferred site be down.

How does NWP change our approach to flight planning?

Early in the article, I commented on the historically strained relationship between forecasters and pilot. Arising from this, some pilots believe that looking at forecasts before the day of flight is a waste of time and that one should wake up in the morning and go on the basis of METARs and TAFs. There is nothing wrong in such an approach but the availability and accuracy of NWP products mean that we have an alternative should we so desire. It is also worth noting that such attitudes often go back to training 20 or 30 years ago, since which time the forecasting skill of NWP has improved by a useful 2 or 3 days.

Light piston IFR flying can involve some quite complex decision making, especially if it is a “got to be there flight”, or involves customs pre-notification, or limited opening hours, or any of the other issues that can affect a flight. I find it easiest to build a picture of the weather for the flight a day or two in advance and then formulate my plan. All I do on the day of flight is to use the available forecast to confirm whether my picture is right. It nearly always is. Interestingly, when delivering the PPL/IR Europe IFR weather seminar, I find that when I give participants less time for assessing a given scenario, the more I get the response to cancel. It is perfectly reasonable that if pushed for time, a pilot should err on the safe side. Therefore, I propose that for some pilots, building a picture in advance could well lead to a higher dispatch rate. There is also a philosophical point. Once upon a time there used to be an official forecaster, who would act in an advisory role. I have had a forecaster say to me “just leave it a couple of hours, and you’ll be fine”. Time has moved on and there is no longer public money available for such a briefing service for private pilots. Regulation has moved on as well. The regulator is looking to make a lighter touch and provide measures that depend on the extent to which the persons affected by the risks involved in the operation, are able to assess and exercise control over those risks. In my view, this also signals a shift from the idea of “official” weather sources towards the pilot becoming the forecaster. The wide range of NWP products certainly makes this possible.

What about the future?

We should expect a relentless increase in the skill of NWP forecasts. This could mean greater accuracy for the same time range or an increased time range for the same accuracy. However, it is unlikely there will be a step change. The rate of improvement of range of about a day per decade should continue, and probably speed up a little. Improvements will come from ever-finer gridcell resolution, and this will bring also better forecasting performance of smaller weather phenomena like thunderstorm cells. Increasing the number of observations fed into the model will also improve accuracy. Over time, algorithms will improve by the process of machine learning, and new algorithms will be developed, such as those needed for forecasting icing. All of this will require time, budget and continuing exponential increases in computing power. We should also expect to see continuous improvements in the graphic presentation of forecasts. In particular, I would expect to see soon a 3D “fly through” presentation of a forecast. Referring to the cautionary warning earlier, I would also expect graphic presentation to improve faster than forecasting skill, so take care!

To conclude, there is one important idea. At the start, I said that NWP is moving faster than the training industry. It is also moving faster than the printed word, so whilst the spirit of this article may continue to ring true, the details will soon become obsolete. Some of my own thoughts included in this article may also be out of date. Help us all by posting your ideas and updates on the PPL/IR Europe Forum.

Events

More Events »

Latest Forum Topics

  • Please login to view posts.

More Topics »