11 August 2015

The Real Reason U.S. Weather Models Missed The Forecast on Hurricane Sandy

Posted by Dan Satterfield


Hurricane Sandy was Forecast by U.S. Models to turn out to sea, while the European models forecasted a hit near New Jersey. Forecasters like me knew the Euro was most likely correct (and it was), but we were wrong about why.

Among my fellow meteorologists, the talk after Hurricane Sandy was how accurate the european ECMWF model was compared to the U.S. Global Forecast System (GFS) model. Fortunately, most of us who were forecasting Sandy, knew that the euro was the better choice from previous experience, but it turns out that the reason we thought it did a better job was wrong. We thought it was because the Euro models initialization of the atmosphere was better than the GFS. In other words, the euro’s starting point was more accurate, and the model had the advantage right from the beginning. Nick Basill (at the Univ. of Utah) has shown that we were all wrong. The paper was published last year but I just ran across it this week, and i think most forecasters may be unaware of it.

First some Background

Wind flow at 700 hectopascals. From here.

Wind flow at 700 hectopascals. From here.

More on why the euro was best in a moment, but first some basics about how we forecast the weather. Numerical weather models have to be given a starting point of what the atmosphere is doing now, but there is one problem with this.

We don’t know.

Oh, we have a very good idea, but most of the planet is ocean, with sparse weather observations, and the same is true over land in most parts of the world. So we interpolate horizontally and vertically to give an accurate starting point to the models. The euro model does this in a much more sophisticated way than U.S. models, because they have faster computers.

The War On Science

The anti-science posture of many of our so-called political leaders is actually having an impact on many aspects of how science is done in this country. The effects are widespread, and run from NASA, to weather prediction, to a serious deficiency in weather satellites. While NOAA now has new computers going online, we’re going to have a tough time catching up, because Europe,China and Japan are not sitting still. More so, they aren’t sitting around arguing over stickers in Biology books saying that natural selection is just a theory, or fighting over teaching accurate climate science to high school students. The result is that they have better weather models, faster computers and much more sophisticated weather satellites that can feed data into those models. 

It’s The Physics Stupid

That was Dr. Ryan Maue’s reaction to the findings published in Geophysical Research Letters (The paper is open access and you can read it here). It turns out that the reason the GFS was wrong is in the physics of the cumulus parameterization. Numerical models have to parameterize processes in the atmosphere that occur at scales that are far lower than the grid size. The euro runs on a smaller grid than the GFS but both have to deal with the effects of cumulus clouds using a separate scheme that estimates the effects on the temperature and moisture fields in each grid box. When the GFS model was given a cumulus parameterization like the euro model, it came much closer to a more correct forecast.

The abstract of the paper is here:


The extremely damaging Hurricane Sandy (2012) is noteworthy for the significant track bifurcation among several forecast models approximately 6–7 days before landfall. The operational versions of the European Center for Medium-Range Weather Forecasting (ECMWF) and Global Forecast System (GFS) models exemplify this difference over their runs early in Sandy’s life cycle, as the former routinely forecast a storm which would make a U.S. landfall, while the latter routinely forecast a track toward the central North Atlantic. This was also generally true of their respective ensemble members. This paper demonstrates that these differences were caused not by resolution or initial condition differences but rather due almost exclusively to choice of cumulus parameterization. Simulations performed with the Weather Research and Forecasting model using an ECMWF-like cumulus parameterization in conjunction with GFS initial conditions yield forecasts whose accuracy is similar to that of ECMWF forecasts at extended time ranges (up to 1 week before landfall).