as it describes different possible scenarios of the future weather.
Prediction of the future - life is more unpredictable than weather
You know nothing about what is actually coming up on you in the future. You set goals and assume different things might happen in your life, but do you really know for certain that it will happen? Probably not, as it is the future we are talking about.
The same goes for weather in the future. We - the weather forecasters - can analyse and make our guesses – qualified guesses. Regardless of this, we actually don’t know the truth.
Our Numerical Weather Prediction (NWP) models do help us, as they most often are close to the truth. We can confidently say that our predictions are closer to the truth, than what you as a child predicted you would work as – after all, we haven’t seen all those spidermen, princesses and football players that we talked about in preschool.
It would be unfair to compare you as a child, with a grownup experienced weather forecaster. Still, if we talk about today, do you really know if you will get home in the exact time you think you will? Maybe you will get a flat tire, stay late at work to solve some crisis or perhaps decide to join your colleagues for dinner. We would argue that life is more unpredictable than weather, at least for the details just a few days ahead of us.
However, the impact of errors in weather predictions can have extensive effects and delay projects. Why we can predict weather better than our lives, is because we have certain tools for it. One of the crucial tools for making a high-accuracy forecast, is the ensemble forecast. This tool is especially crucial the further into the future you try to predict.
Ensemble forecasts? What are we talking about? Basically, it is a collection of several different weather forecasts – either based on different models and/or different initial conditions.
Before going through why and what this means, you might need a little insight in how these NWP models work. NWP models are a selection of equations with the aim of describing the development and motion in the atmosphere. One problem arise as meteorology is not a perfect science, so no set of equations are perfect. There is a lot of different ways of making the NWP models - this makes the models to differ from each other. This means that if comparing two weather forecasts, they will not give you the exact same weather prediction.
Another of the main challenges come from the initial state, as these models need to know the “right now” in order to calculate anything of the future. This initial observation data come from measuring stations worldwide, satellites, airplanes etc. The observation equipment isn’t perfect and may give you values with some error. Even when trying to adjust the values, it may not give the perfect representation – and on top of this, there are way too few measuring stations to cover every cubic meter of the atmosphere. This means that parts of the initial states are partly based on whatever is going on many kilometers away – but numerically adjusted. It is almost as assuming that you and your neighbor always will wear the same clothes but in different colors…
To make this a bit more visual - if you would like to build a wall, as the one above in the figure, but you want it to be completely vertical. You can see that it is the bottom bricks – the initial state - that make the structure tilt more and more with height. If you continue to build on this wall, it will eventually fall.
This is sort of what is happening with weather forecasts. You also have different models that have small errors within, visually we could represent it as the hole in the wall – which also makes it more unstable. Try to make this wall vertical if you still must have the toy car in the base of it. Maybe you try to improve the stability of the wall by moving the toy car slightly back and forth a few times. This is also sort of what is done with ensemble forecasting - by adjusting initial conditions a little and then run the model for a set of different initial conditions. All these model runs can be plotted on top of each other for a certain parameter that you are interested in. Alternatively, you can compare different models for the same location and time in the same manner. This means that you now have a number of different possible weather developments that you can evaluate on.
How to interpret ensemble forecast data
An example of an ensemble forecast is shown in the above figure. The data is from the GFS model which is run with different initial states - the output parameter is the mean wind in 10m elevation in Copenhagen. The lines represent the model output based on several different initial conditions, as well as one for the average for all model runs (black) and one for a high-resolution operational model run (green). The “operational” is the one you would see if you checked the weather online/app for Copenhagen (if the forecast you look at is based on the GFS model).
You can see a rather clear pattern within the first few days (like winds between 5 and 7m/s on May 10th), but the further you go ahead in time, the more does the model runs disagree. Just as that unstable brick wall you were building, would tilt more and more the higher you would build it.
Even when the model runs are not aligned, it can tell something about what is most likely to happen. It builds on probability - the more models and/or model runs that are pointing in the same direction, the more likely it is to occur. Here is also where the meteorological experience is a part of this evaluation.
If you are one of our clients, you may have seen below figure before. This is almost the same as in the figure (GFS) above, but just presented in a different way, with a different NWP model and for a slightly different location. Here the shaded grey colors with the white line inside are the ECMWF ensemble forecast data.
Now you can’t see every model run as in the previous figure. It is instead the mean value (called average before), white line, then the 25th to 75th percentile, dark grey, and the 10th to 90th percentile, light grey. The mean and these percentiles give indications on what is most likely.
On top of this, you have a black line, which is for the first two days, a high-resolution (grid spacing of 2.5km) regional model. This is not ensemble data, but a single model run. However, it is very convenient to have this model on top of an ensemble, as it at least for an experienced user, can give further information.
We know that sometimes our regional model is not that good, even if it is of very high resolution and quality. We often know in which meteorological occasions to doubt our model somewhat, and when having the ensemble behind the black dotted line, we can justify our doubts and help us in making a better forecast evaluation for a specific site at sea than the model does.
If you look at Tuesday, May 8th and Wednesday, May 9th, you might think that the model is totally wrong where the wind is peaking (black dotted line) - if you compare it to the gray shading. However, this is probably not the case, since the model are forecasting for different sized boxes (called grids). The regional model has smaller boxes being forecasted, while ECMWF is a global forecast model that have larger boxes (presently a grid spacing of 9km).
If you have a large box of different colored candy and make a mean value of red candies in the whole box, you get one value. But if you make a mean value of the separate four corners, where one of the corners contain a lot of red candies, it will not give you the same picture. If you like red candy your now know where to put your hand if you want to increase your chance of getting a red one to eat. This is as the peaks in wind - it might not be forecast by the ensemble model because it makes a mean of the larger global box system instead of the smaller regional one.
We hope that you now know something about how we take some of the uncertainty out of the weather forecasts.