Disentangling forecast errors

By Oscar Martínez-Alvarado

In a previous blog entry I discussed the factors contributing to errors in weather forecasts. These factors were sensitivity to initial conditions, incomplete knowledge of the atmosphere’s state and incomplete representation of the laws of physics governing the atmosphere’s evolution. At this point I would like to make clear that the presence of forecast error does not render a forecast useless. Indeed, being able to approximately predict what the atmosphere will do in the hours and days ahead is an enormous human achievement. It is only due to this ability that we can limit the impacts of storms that would otherwise be catastrophically damaging. An example of a storm that was forecast sufficiently well is the ‘St Jude’s Day’ storm on 28 October 2013. Thus, the presence of forecast error only indicates that our ability to forecast the weather can be improved further by understanding that error and hopefully reducing it. Scientific endeavours always work in this way. For example, an accurate description of the planets’ orbits could not have been achieved by Newton without a long chain of people before him devoting time to the observation of the sky that started even before the ancient Greek philosophers.

Coming back to our topic, to understand forecast error the first thing to do is to disentangle the contributions for the different factors that contribute to it. The three factors occur simultaneously and therefore separating them is not an easy task, especially when these factors all start being very small. It does not occur often, but from time to time the forecasts produced by different agencies deviate in similar ways from what actually happens. These common discrepancies can be used to explore the causes of such failures.

As an example I will use the case of forecasts of Rossby waves. Rossby waves constitute the trough/ridge pattern along the jet stream at tropopause level (about 10 km above sea levels in midlatitudes) and are key elements of the atmospheric circulation. Recent work developed at the Department of Meteorology (see Gray et al. 2014) has shown that, among other features, the amplitude of Rossby waves decrease with forecast length. What this means is that a Rossby wave will appear weaker in a forecast made five days ahead than in a forecast made one day ahead. This same behaviour can be found when the forecasts are produced by different models, which points to a common forecast error. Once we have identified a common forecast error we can then investigate further into this behaviour using case studies. One such case study appears in Figure 1, which shows that the error in the amplitude of the Rossby wave on 24 January arose from a very small cyclone off the coast of Florida that the forecast failed to represent at the right time.

2015 09 07 Oscar M-A Figure1


Figure 1. Comparison between the actual evolution of the atmosphere and a forecast valid on 22 January 2011 (upper row) and 24 January 2011. Grey lines represent the coastline; thin black lines represent sea level pressure; bold black lines represent Rossby waves around 10 km above sea level. The blue rectangles show important differences: an incipient cyclone off the coast of Florida not present in the forecast on 22 January 2011 (upper row) and a large-amplitude Rossby wave that has a smaller amplitude in the forecast on 24 January 2011 (lower row). The cyclone on the 24 January was also deeper in reality (966 hPa) than in the forecast (978 hPa).

We analysed this case study in a recent paper (see Martínez-Alvarado et al. 2015) to show that the connection between the error in the cyclone development and that in the Rossby wave was given by the warm conveyor belt, which is one of the main air streams driven by the cyclone (see, for example, Figure 1 in this blog entry). The warm conveyor belt is a cloudy feature in which large amounts of water condense, heating the atmosphere and changing wind motion. In the paper, we also showed that there were differences in heating between reality and the model and that those differences were the likely culprit for forecast error.

Gray, S. L., Dunning, C. M., Methven, J., Masato, G. and Chagnon, J. M., 2014. Systematic model forecast error in Rossby wave structure. Geophysical Research Letters, 41, pp. 2979-2987. http://onlinelibrary.wiley.com/doi/10.1002/2014GL059282/abstract

Martínez-Alvarado, O., Madonna, E., Gray, S. L. and Joos, H., 2015. A route to systematic error in forecasts of Rossby waves. Quarterly Journal of the Royal Meteorological Society. http://onlinelibrary.wiley.com/doi/10.1002/qj.2645/abstract


This entry was posted in Numerical modelling, University of Reading, Weather forecasting and tagged . Bookmark the permalink.