Why are short term weather forecasts still so iffy?

[goenetix]

@astro

From my layman’s perspective prediction accuracy doesn’t seem to have increased (real world) a whit in the last 10 years in what is being delivered to me as a consumer of weather projections.

So, how can we tell a bad meteorologist from a good meteorologist? I can’t imagine how a weatherman asks his boss for a raise: “Weeeeell, the accuracy of my weather forecasts is now solid 40% compared to miserable 32% when I was a rookie five years ago - so I really deserve that raise”

[Shalmanese]

Is the person reading out the weather on the news actually a meterologist? I always thought they got their data from some centralised source and just parroted out the results. In fact, come to think of it, it would be a perfect job to outsource.

[Mathochist]

@Shalmanese

Is the person reading out the weather on the news actually a meterologist? I always thought they got their data from some centralised source and just parroted out the results. In fact, come to think of it, it would be a perfect job to outsource.

Sometimes it is, and sometimes it isn’t. I remember the weather guy on the news my parents watched when I was a kid was a real meteorologist since it turned out he knew a friend of my parents who was a professor of meteorology.

[astro]

@Mathochist

Five orders of magnitude in a week is just the first example off the top of my head. There are similar problems in the shorter term, partially because the longer term is made up of the shorter term…

Basically, mathematicians have done all they really can do with it. The problem is as solved as it’s going to get.

Well, not all the mathematicans are sitting on their sliderules.

Computing the weather

When modelling the weather in any location, it is important to consider conditions occurring elsewhere in the world. “The weather is highly sensitive,” explained mathematician Martin Gander. “Initial conditions become erratic over time.” A small, virtually undetectable change today could greatly affect the weather days later. “The butterfly effect is a commonly used analogy,” Gander continued. “The flap of a butterfly’s wings in Japan could set off a snowstorm in Montreal, it’s that sensitive.”

Gander, who has developed mathematical algorithms to describe countless processes, from chicken cooking in a microwave oven to the movement of frigid air through his apartment during a bitter Montreal winter, admits that incorporating the flight of Asian butterflies into a model of Canadian weather may be a little unrealistic. However, the benefit of including larger scale elements, such as an Asian monsoon, is obvious. Unfortunately, the mathematical algorithms currently used are incapable of even considering events of this magnitude when calculating the weather in Canada.

“Our current weather prediction algorithms involve math found in 30-year-old textbooks,” noted Gander. “It’s like going to a scrapyard to buy a computer.” Mathematics has advanced at lightning pace since the '70s. Many consider the evolution of mathematical algorithms to be on par with that of the computer. Unfortunately, weather prediction math was not updated to run on modern computers, so our forecasting ability remained as dated as tie-dye and disco. According to Gander, a mathematical make-over in order to incorporate global weather elements into local weather predictions was long overdue.

The principle Gander used is as old as it is simple. “Divide et impera,” commented Gander. “Divide and conquer.” In order to predict weather to a suitable resolution, the world must be divided into manageable pieces; the weather in each geographic cell can then be resolved by an individual computer processor. Naturally, the weather in each cell is influenced by the weather in another. The significant weakness of the old algorithms was the treatment of each geographic cell as an independent entity.

[Mathochist]

@astro

Well, not all the mathematicans are sitting on their sliderules.

Computing the weather

Again, this is really marginal improvement. You simply can’t get any better in air pressure than an error of about 100,000 times your measurement error within two weeks time, and that’s about the best-behaved weather equation there is.

[Viscera]

Sorry for bumping a month-old thread, but I don’t visit here too often, but when I do it’s in spurts. Anyway, I wanted to add some additional thoughts as a resident meteorologist. Some very good points were raised and cited.

The main reason that medium range forecasts (past a couple of days) is not as accurate as one may hope is due to relative lack of data, and the relative primitiveness (is that a word?) of the numerical models we use. Yes, some of the models we now use are very high resolution (with grid points less than 10 km on many of them), but that doesn’t necessarily help at some point. Our upper-air observing stations are on average about 300 miles apart. These are the stations that launch balloons twice a day for a full cross-sectional sampling of the atmosphere. Other forms of data are helping out, which include satellite-based samplings (huge positive effect of being able to sample tons of different points).

Anyway, even if the data problem is solved, even with a fine-scale model, lets assume say 10km. The model will not be able to capture any feature that is less than 20km in diameter. Local effects such as rain bands, individual thunderstorms, lake, can’t be modelled at this point in time. Granted, through parameterization (which I’m not sure of the details of, but can find out if anyone is really interested), we can make the model output better, but the model still can’t do everything.

Finally, as has been mentioned, the chaos of the atmosphere is the last reason. There are so many things happening on the order of tens of meters (from local fluxes, turbulence, gravity waves, etc. that are extremely difficult to model, and must be parameterized to try and capture them). Take these out 3-4 days, and it’s no wonder that local effects can seem way off.

All that said, our experience with large-scale weather patterns is very good. We have the same accuracy now at 6-7 days that we did maybe 15 years ago at 2-3 days (I’m sure these numbers aren’t exact, I’m a bit rushed for time, but I can find something if needed). We are now forecasting hurricanes with the same accuracy at 4-5 days that we did at 2-3 days 10 years ago. As computer power continues to progress exponentially, we will continue to see better forecasts, and as data retrieval techniques combined with higher resolution modelling get better, we will see much better forecasting of smaller-scale effects.

[Mathochist]

@Viscera

Anyway, even if the data problem is solved, even with a fine-scale model, lets assume say 10km. The model will not be able to capture any feature that is less than 20km in diameter. Local effects such as rain bands, individual thunderstorms, lake, can’t be modelled at this point in time. Granted, through parameterization (which I’m not sure of the details of, but can find out if anyone is really interested), we can make the model output better, but the model still can’t do everything.

Finally, as has been mentioned, the chaos of the atmosphere is the last reason. There are so many things happening on the order of tens of meters (from local fluxes, turbulence, gravity waves, etc. that are extremely difficult to model, and must be parameterized to try and capture them). Take these out 3-4 days, and it’s no wonder that local effects can seem way off.

Just a comment on chaos, for those with only a pop-culture knowledge. This is true no matter how fine the scale is.

If the current technology has a resolution of 10km, then smaller features than that can’t be seen. If you make an assumption (and you have to) about how it behaves in there, your error will be smaller than 10km. This will blow up relatively quickly and completely throw off your calculations no matter how accurate your model is.

Now, it’s easy to see how being off in a 10km sample outside New York City will make it difficult to see what’s going to happen next week in Boston, but when you really realize that being off in a single 10km sample in Hollywood throws off all the calculations for the entire United States within a relatively short period of time you really get an idea what “chaos” means.