Early Numerical Weather Models
from Gerhard Holtkamp,
07. February 2010, 00:26
A look back into the past can help us understand what we have today. Here are some memories about numerical weather forecasting four decades ago...
Maybe it was the annoying accuracy that the poor weather we keep having this European winter was predicted which made me look in the attic for an old brochure from the year 1972. This brochure gave an overview of the numerical weather models in operational use at the German Weather Service and I kept it as a souvenir to remind me of my time there at the beginning of my professional life.
Attempts at numerical weather forecasting started almost as soon as electronic computers became available in 1950 by non other than John von Neumann together with J. Charney and R. Fjörtoft. They did so by treating the highly complex physics of a three dimensional atmosphere as the surface of an incompressible fluid.
The height of this surface would model high or low pressure systems and the movement of this surface indicated horizontal wind speeds. There was no temperature, radiation or moisture present - all the things we normally complain about when we talk about the weather.
This drastic simplification was due to the slow speed of those early computers. If the results of numerical weather calculations are to be of any operational use then the time to come up with a 24 hour prediction must not exceed about an hour. So a pattern was set right at the beginning which continues until today: The complexity of numerical weather models expands to fill the computing power available.
With some minor modifications this simple so called barotropic model continued to be in operational use well into the 1970s even though more sophisticated models were already employed at that time.
While the basic equations underlying the barotropic model came straight from introductory fluid physics textbooks meteorologists had to come up with some specific additions to make them of any practical use.
The physics of the model would still allow short period sound and gravity waves. To avoid numerical instabilities when integrating the equations time steps could not exceed a few seconds which was unacceptable for the slow computers of that age. But meteorologists knew that those waves would have no influence on the short term behaviour of large scale pressure systems which was all that the barotropic model could handle anyway.
So meteorologists did what physicists, engineers and photographers love to do: They applied some filters. Thus getting rid of those unwanted waves the integration time could be increased one hundred fold and practical numerical weather predictions had become possible.
There was also the question of exactly which surface to use in this two dimensional model. Knowing that the highly simplified assumptions would best be approximated at the 500 mb pressure level (which corresponds to roughly 5000 m altitude) it was this surface which was chosen for the calculations.
A few more details (like an attempt to also model the form of the Earth surface and friction) were added in some versions and a number of mathematical transformations were done on the original equations to make them more suitable for numerical integration.
Once the simple barotropic model was introduced into operational weather forecasting the accuracy of predicting the wind (at the 500 mb level) increased by 20% as did the prediction of the movement of large scale pressure systems. Still the many simplifications of the model led to large errors once you passed the 48 or 72 hour mark.
At the beginning of the 1970s a more advanced model was in operational use. The so called baroclinic model is a three dimensional model and needed an order of magnitude more computing power. It did include temperature but radiation, clouds etc. were still missing. In essence it treated the atmosphere as an ideal gas. The spacing of the horizontal grid points was about 400 km and the vertical layers were spaced 2-3 km apart. Thus countries like Germany, Britain or France were represented by just about 3 or 4 (horizontal) grid points!
Mountains were included but had to be grossly simplified. As it turned out the Alps were modelled best if taken with an elevation of just 700 m! Numerical weather predictions were run by weather services in different countries. Although the model itself was the same there are always some parameters to tune and each country did it slightly differently so that on average it came out best for that region.
The German Weather Service regularly exchanged their numerical predictions with those from the British and American service. The British forecasts tended to be more on the pessimistic side of low pressure systems which was not surprising seeing that Britain is a more exposed maritime nation.
The reason the simple barotropic model was still kept during those years was that only the larger of our two mainframe computers could handle the more sophisticated baroclinic model in a timely fashion. If that computer broke down you still had the simple model to work with.
A small third computer served as a communication gateway by storing the incoming weather reports on magnetic tape. Apart from a few trunc lines between the major weather services most weather reports went via slow teletype lines. In fact if the communication computer broke down the data had to be read in via punched tapes from the teletype writers!
There were no direct connections between our computers as would be the norm now so the operators had to keep mounting and unmounting a myriad of magnetic tapes to transfer all the data from the communication computer to the prediction computer and on to the four large mechanical plotters which had to draw the multitude of weather charts needed.
Soon we would get a new top-of-the-line computer 20 times faster than the old one. This allowed the grid size of the model to be cut in half yielding a more accurate forecast. The ever increasing computing power in the following years also meant that more complex weather models could be introduced including radiation, humidity and clouds. The numerical drought was over. At long last there was some rain falling out of the models!
But even the best numerical models are only as good as the data you feed in. Here was another major development happening at the beginning of the 1970s. Weather satellites had advanced from mere cloud picture taking to proper atmospheric sounders providing weather data globally even for areas like the oceans for which there had been a severe lack of data before.
Weather satellites also created a new problem. Up to that time data at weather stations all over the world had been gathered simultaneously at fixed times. But satellite data came in asynchronously for different places. To get all the data into the fixed (space and time) grid points needed by numerical weather models something called Four Dimensional Weather Analysis was developed which was a hot research topic at the time.
The intervening decades since have seen dramatic advances in computing power, satellite technology and more detailed modelling of atmospheric, oceanographic etc. physics and their interdependence. Numerical weather predictions are no longer restricted to large scale phenomena but are getting better all the time even at small scale local events.
But there is a limit to all this. About four decades ago something else happened: The notion of chaotic dynamic systems slowly became acknowledged by ever more scientists. We will never be able to predict the weather with arbitrary accuracy long into the future. And that's good. How else could we start a conversation if we didn't have the weather to discuss?
Maybe it was the annoying accuracy that the poor weather we keep having this European winter was predicted which made me look in the attic for an old brochure from the year 1972. This brochure gave an overview of the numerical weather models in operational use at the German Weather Service and I kept it as a souvenir to remind me of my time there at the beginning of my professional life.
Attempts at numerical weather forecasting started almost as soon as electronic computers became available in 1950 by non other than John von Neumann together with J. Charney and R. Fjörtoft. They did so by treating the highly complex physics of a three dimensional atmosphere as the surface of an incompressible fluid.
The height of this surface would model high or low pressure systems and the movement of this surface indicated horizontal wind speeds. There was no temperature, radiation or moisture present - all the things we normally complain about when we talk about the weather.
This drastic simplification was due to the slow speed of those early computers. If the results of numerical weather calculations are to be of any operational use then the time to come up with a 24 hour prediction must not exceed about an hour. So a pattern was set right at the beginning which continues until today: The complexity of numerical weather models expands to fill the computing power available.
With some minor modifications this simple so called barotropic model continued to be in operational use well into the 1970s even though more sophisticated models were already employed at that time.
While the basic equations underlying the barotropic model came straight from introductory fluid physics textbooks meteorologists had to come up with some specific additions to make them of any practical use.
The physics of the model would still allow short period sound and gravity waves. To avoid numerical instabilities when integrating the equations time steps could not exceed a few seconds which was unacceptable for the slow computers of that age. But meteorologists knew that those waves would have no influence on the short term behaviour of large scale pressure systems which was all that the barotropic model could handle anyway.
So meteorologists did what physicists, engineers and photographers love to do: They applied some filters. Thus getting rid of those unwanted waves the integration time could be increased one hundred fold and practical numerical weather predictions had become possible.
There was also the question of exactly which surface to use in this two dimensional model. Knowing that the highly simplified assumptions would best be approximated at the 500 mb pressure level (which corresponds to roughly 5000 m altitude) it was this surface which was chosen for the calculations.
A few more details (like an attempt to also model the form of the Earth surface and friction) were added in some versions and a number of mathematical transformations were done on the original equations to make them more suitable for numerical integration.
Once the simple barotropic model was introduced into operational weather forecasting the accuracy of predicting the wind (at the 500 mb level) increased by 20% as did the prediction of the movement of large scale pressure systems. Still the many simplifications of the model led to large errors once you passed the 48 or 72 hour mark.
At the beginning of the 1970s a more advanced model was in operational use. The so called baroclinic model is a three dimensional model and needed an order of magnitude more computing power. It did include temperature but radiation, clouds etc. were still missing. In essence it treated the atmosphere as an ideal gas. The spacing of the horizontal grid points was about 400 km and the vertical layers were spaced 2-3 km apart. Thus countries like Germany, Britain or France were represented by just about 3 or 4 (horizontal) grid points!
Mountains were included but had to be grossly simplified. As it turned out the Alps were modelled best if taken with an elevation of just 700 m! Numerical weather predictions were run by weather services in different countries. Although the model itself was the same there are always some parameters to tune and each country did it slightly differently so that on average it came out best for that region.
The German Weather Service regularly exchanged their numerical predictions with those from the British and American service. The British forecasts tended to be more on the pessimistic side of low pressure systems which was not surprising seeing that Britain is a more exposed maritime nation.
The reason the simple barotropic model was still kept during those years was that only the larger of our two mainframe computers could handle the more sophisticated baroclinic model in a timely fashion. If that computer broke down you still had the simple model to work with.
A small third computer served as a communication gateway by storing the incoming weather reports on magnetic tape. Apart from a few trunc lines between the major weather services most weather reports went via slow teletype lines. In fact if the communication computer broke down the data had to be read in via punched tapes from the teletype writers!
There were no direct connections between our computers as would be the norm now so the operators had to keep mounting and unmounting a myriad of magnetic tapes to transfer all the data from the communication computer to the prediction computer and on to the four large mechanical plotters which had to draw the multitude of weather charts needed.
Soon we would get a new top-of-the-line computer 20 times faster than the old one. This allowed the grid size of the model to be cut in half yielding a more accurate forecast. The ever increasing computing power in the following years also meant that more complex weather models could be introduced including radiation, humidity and clouds. The numerical drought was over. At long last there was some rain falling out of the models!
But even the best numerical models are only as good as the data you feed in. Here was another major development happening at the beginning of the 1970s. Weather satellites had advanced from mere cloud picture taking to proper atmospheric sounders providing weather data globally even for areas like the oceans for which there had been a severe lack of data before.
Weather satellites also created a new problem. Up to that time data at weather stations all over the world had been gathered simultaneously at fixed times. But satellite data came in asynchronously for different places. To get all the data into the fixed (space and time) grid points needed by numerical weather models something called Four Dimensional Weather Analysis was developed which was a hot research topic at the time.
The intervening decades since have seen dramatic advances in computing power, satellite technology and more detailed modelling of atmospheric, oceanographic etc. physics and their interdependence. Numerical weather predictions are no longer restricted to large scale phenomena but are getting better all the time even at small scale local events.
But there is a limit to all this. About four decades ago something else happened: The notion of chaotic dynamic systems slowly became acknowledged by ever more scientists. We will never be able to predict the weather with arbitrary accuracy long into the future. And that's good. How else could we start a conversation if we didn't have the weather to discuss?
