Motivation...
DR. ROBERTO BUIZZO: If we go back 25 years ago, we used to have one single forecast issued every day. The problem we had was that we didn't know whether the forecast was going to be accurate or not.
DR. FLORENCE ROBIER: When we take observations of the atmosphere, you know, it's like when you have a thermometer in a room – when if you put another thermometer next to it, you will not necessarily have the same reading. Because nothing is perfect. Which means, there's always uncertainty in what we measure and what we predict.
TOM HAMILL: Ed Lorenz was one of the first to really cast in a mathematical framework with which he was able to demonstrate that if you make a tiny error at the initial time, that tiny error is going to grow into a very substantial error.
DR. FLORENCE ROBIER: We always say that atmospheric weather prediction is an initial value problem. So, we first determine what is the state of the atmosphere now, what is the initial value, the initial conditions, and then we run a model. Traditionally, people were just using one model, typically in one center. They were just doing an analysis of the atmosphere using one model and then getting the forecast for the next few days. And that was very much what we were doing. We were taking the best image of the atmosphere with the observations, the best model we could get, and we had one shot in the future to say what the forecast is.
Ensemble prediction is about running one model several times; and in particular, here at ECMWF, we were trying to change the initial conditions of the forecast.
DR. ROBERTO BUIZZO: For each point for the globe, instead of having just one value of temperature, we had 50 value of temperature at the initial point. The same for wind, the same for pressure.
DR. FLORENCE ROBIER: And this would actually trigger the atmosphere to go in different directions in our forecast.
DR. ROBERTO BUIZZO: Then I can estimate using all of them what's going to be the possible range of forecasts, and this was, for example, extremely helpful in the summer 2003 when Europe experienced extreme hot temperatures. So with our ensemble systems, we're able to predict these extremes two, three weeks ahead.
DR. FLORENCE ROBIER: What we really use a lot is tropical cyclones and the trajectory of tropical cyclones, so there you look at it on a map not on a given point, but you really look at the different trajectories of the tropical cyclones.
TOM HAMILL: You probably have heard of hurricane or superstorm Sandy along the east coast of the United States, and that was a case where the ECMWF prediction system provided the heads-up a day or two in advance of our own US Weather Service prediction system. That was a success story.
DR. FLORENCE ROBIER: So, I think at the beginning, we were trying to really push the forecast where it was going to be sensitive. Then, we started also to incorporate some idea on the uncertainty in the observations. Sort of quantifying actually what comes from the uncertainty in the observations themselves. And, finally, the third step has been to actually quantify what is the uncertainty in the model itself. We know there are some ranges where some parameters could lie, for instance, and so, it's quantifying these as well.
DR. ROBERTO BUIZZO: Now, if I think about the future, we want to make our models better. We know, for example, that coupling the models to the ocean is key, and we have started doing that. We want to improve the initialization of the model to our estimation of state of the atmosphere. Third big area of work is resolution. So, we need more resolution in the system.
DR. FLORENCE ROBIER: We have a very high collaboration goal, that we really think we benefit if we collaborate with other organizations.
TOM HAMILL: Oh, my colleagues and I really think the world of ECMWF. The scientists are really defining the forefront of the operational ensemble predictions, and where the research needs to head to address the remaining deficiencies in ensemble prediction systems.
DR. FLORENCE ROBIER: It's fantastic, extremely motivating to know that you're working in a place where you can really make a difference, and whenever we have a good success in our forecast, and we think we might have saved lives using the science to do that. I can't think of anything better to do.
The video above provides an overview of ensemble forecasting at ECMWF. As technology has advanced, forecasting has improved. NWP models can now create multiple runs for the same forecast time, creating an ensemble. Ensembles provide more confidence in forecasts, at least when runs coincide with one another, and tell the forecaster where and when higher uncertainty exists (where the runs diverge). Thus, ensembles can be used to assess the amount of uncertainty in a model forecast and the range of likely outcomes. In this lesson, you will learn more about ensembles, including how to use them, how to assess the model uncertainty, and how to turn the ensemble into a PDF for application purposes. Read on to learn more!
Lesson Objectives
- Explain what an ensemble is and discuss its strengths and weaknesses.
- Discuss the different sources of model error and how ensembles can be used to assess the resulting forecast uncertainty.
- Turn ensembles into PDFs and calibrate the PDFs.
Newsstand
- Developing a State-of-the-Art Verification Toolkit. Retrieved June 27, 2019
- Scherrer, S.C., Appenzeller, C., Eckert, P., & Cattani, D. (2004, June). Analysis of the Spread–Skill Relations Using the ECMWF Ensemble Prediction System over Europe. Retrieved June 27, 2019.
- Whitaker, J.S. and Loughe, A.F. (1998, December). The Relationship between Ensemble Spread and Ensemble Mean Skill. Retrieved December 7, 2018.
- Peña, M. (2011, March). Ensemble Forecasts and their Verification. Retrieved December 7, 2018.
- Hersbach, H. (2000, October). Decomposition of the Continuous Ranked Probability Score for Ensemble Prediction Systems. Retrieved December 7, 2018.
