When you need a weather forecast the most – like when a hurricane is bearing down on you – is when the predictions seem to fail. Why is that?
As Irma came roaring in from the Caribbean with winds of up to 185 mph, weather services which were able to predict Irma’s birth were unable to predict its behavior. Days before, we didn't know where Irma would hit. Maddening to Florida residents, the path kept moving. They didn't know which side of the state needed to evacuate. Miami braced itself for a full assault, only to have Irma veer to the west.World's Biggest CFD Problem
For the last couple of decades, the ECMWF, the European Union’s global weather model has been the gold standard for accuracy. But it only runs twice a day. And it charges for results.
A US model, the US system, the Global Forecasting System (GFS), runs more frequently - 4 to 6 times a day. It's known to be good at predicting storms. It keeps getting better.
The National Hurricane Center, which runs several models, including the ECMWF and the GFS, to make its “spaghetti charts,” claims increasing accuracy over the years. In 2000, the average miss was 150 miles but now it's 75 miles. A two day prediction is the state of the art.
"A hundred miles in a three day forecast is really good," says Dr. Brian McNoldy, a researcher at the University of Miami quoted in the New York Times. Quick to disagree would be millions of people in Florida, when the 100 miles is the difference between one densely populated coast and the other.
The main reason for the uncertainty is the vast amounts of data over a weather model that is literally the biggest fluid/thermal problem in the world -- and the granularity the scale imposes. A single cell in the global model has sides that are 40 kilometers (25 miles) long, big enough to swallow entire cities leaving insufficient resolution to account for local effects. The widest part of the Florida peninsula would be covered in 6 cells. You will not know what will happen in your neighborhood, much less to your house.
Weather models also have to deal with ridiculous amounts of input, including radar images and satellite data.
In a hurricane, not only are parts in motion, but they are travelling fast. The time slice must be very thin, so more calculations per hour are needed. This is in contrast to predictions of less volatile weather conditions and forecasts that are accurate for many days in advance. You can tell on Monday whether the sun will be out for your hike on the weekend, for example.
We're Getting Better
Weather researchers insist hurricane prediction is getting better all the time. A model's accuracy can be measured in the difference between where it hits land and and how fast its winds are when it hits, compared to the predictions a few days out. Predictions of Hurricane Harvey's (which flooded Houston) landfall varied over a much larger span of coastline than did Irma predictions, but as the Texas coastline is much less populated than Florida's, its uncertainty was less of an issue.
The weather prediction can get a lot better, says Jeff Masters, director of meteorology at the Weather Underground. He cites a 2007 recommendation that hurricane research be increased to $300 million a year, presumably leading to more accurate modeling. While small in proportion to the economic devastation of a major hurricane hitting a big metro, this is 10 times the amount of research currently being done in the United States.
With hurricanes getting stronger (this is first time hurricanes of Harvey and Irma's intensity have hit so close together), perhaps increased research could be seen as the prevention that is less expensive than the cure.