At times during Harvey, the European model outperformed humans – News » At times during Harvey, the European model outperformed humans

At times during Harvey, the European model outperformed humans

Admin Sep 05,2017

Android One just got a hardware upgrade courtesy of Xiaomi.

As a resident of Houston and a meteorologist, I closely tracked the development or Hurricane Harvey for much of the month of August. As part of that, I watched every new cycle of forecast model runs to determine which one handled the track forecast best.

For a long time, the big question with Harvey concerned what the storm would do after it moved inland into the central Texas coast late on Friday, August 25. Would it stall for one or two days and then move south? Would it stall for three or four days? Or would it slowly drift to the east-southeast, out into the Gulf of Mexico and then move up the coast?

Each of these particular solutions mattered for Texas, because whichever area lay to the northeast of the storm’s center would receive the heaviest rains. Ultimately the rain bullseye fell along a corridor along Interstate 10 from Houston to Beaumont, where more than 45 inches of rain fell during a period of about four days that led to catastrophic flooding.

Euro, again

This considerable uncertainty about Harvey’s movement was due to the utter lack of steering currents for Harvey once the storm moved inland and bumped up against a large area of high pressure draped over the southwestern United States. By this point it was essentially a marble, rolling across a flat table.

Not surprisingly, it was the European forecast model that first sniffed out the storm’s actual track. As early as Thursday, August 24, the model forecast a move inland near Victoria, a stalling out over the weekend, and a southeastern drift that brought the storm back over the Gulf of Mexico by Sunday or Monday. This is very close to what actually unfolded over the next five days.

The European model is the best forecasting system in the world for several reasons, most notably because the European modeling center has invested heavily in their model. It has the most advanced computer hardware and has devised the best system to assimilate real-time meteorological observations into its model for future runs. This means the model runs start with the most accurate initial conditions.

Digging into the data

After the storm made its final landfall and moved into the Louisiana coast, I wondered if my casual observation of the European model’s performance matched the data. I reached out to Brian Tang, an atmospheric scientist at the University of Albany who keeps real-time statistics on model performance. Here is the data on Hurricane Harvey and the model performance. Tang confirmed that, overall, the European model had shone brightly during Harvey.

For our purposes, the best tab to look at on his site is “mean absolute error” in forecast track. In Tang’s data for this error, there are several clusters of models worth considering. There are the main global forecast models (North American GFS, Canadian, United Kingdom, and European), NOAA’s primary hurricane-specific models (the HWRF and the new, experimental HMON), and finally a consensus model that averages the track forecasts of other models (TVCN) and the official forecast from the National Hurricane Center.

The numbers at the top of the bar graph represent the number of data points for each model (the European model, for example, has about half as many as the North American model because it runs every 12 hours instead of every six hours). Lower errors, obviously, are better.

Mean absolute error for various forecast models and Hurricane Harvey.
Mean absolute error for various forecast models and Hurricane Harvey.
Brian Tang

When looking at the data, we see several interesting trends. Among the global models, through a 72-hour forecast, the UK model is generally a little bit more accurate than the European model. However at longer time frames, four- and five-day forecasts, the European model shows its class. In fact, the four-day forecasts of the European model were more accurate than the consensus model or even the human forecasters at the National Hurricane Center who weigh all of this data to make their five-day forecasts and draw the cone of uncertainty.

Also of note? NOAA’s new hurricane model, the HMON, performed terribly. I recall during the height of the storm that this model, instead of pulling Harvey toward the Houston area, moved the storm southwestward, into Mexico. This scenario would have spared the fourth largest city in the United States from catastrophic rainfall. Alas, the HMON model’s average error at four days was a staggering 630 nautical miles, compared to about 170 nautical miles for the European model. Oops.