Re: Earl; Real-Time; Overnight
Posted by Chris in Tampa on 9/2/2010, 3:42 am
Now I'm confused, lol. The product description that I wrote that gives me nightmares thinking about it didn't clear it up?
http://www.tropicalatlantic.com/models/data.cgi?page=modelerror
Lol. Best thing on that page is the table. Everything else is a rough draft that I will revisit again many times when I can coherently try to explain it and should not be read until then probably.

It is a complex feature. The table on the page above kind of explains it well. (It's confusing, but it kind of makes sense after awhile, I think.) Basically, the "single run" error type, which I can't find a good name for, is just a single run's worth of calculations.

For this current position, a particular model 6 hours ago said the storm would be in this position. That is 6 hour error for an early cycle model.  For position error, I simply calculate the distance from the current storm position to the storm position the model forecast 6 hours ago. And I do that for every position available. If the model made a 120 hour forecast 120 hours ago, I compare that forecast coordinate to the current storm position. (That would be 120 hour error.) Late cycle models make things a little (a lot) more confusing since they are actually initialized 6 hours prior to the current storm position. That is where the table can help explain things kind of.

So I do a set of these calculations, which I call the "single run" error type, every 6 hours when model data is released. For the "average" error type, I add up all the error.

If I have 10 "single run" error values for a certain model for the error hour 6, I add them up and divide by 10.

It's confusing to explain. The chart might help. Think of the chart being done four times a day. Any model error data that cannot be calculated means nothing goes in that square of the chart. When it is time to calculate the average, I add up all the 6 hour error values and divide by the number of instances something appeared in that table square. That gives me the average.

This is not to be confused with the average row that appears on a page which is the average of all the values on the page for that column. Then the heat map gives green colors if something is better than average and red for worse than average.

And finally, the "non-homogeneous comparison" thing. Early cycle and late cycle models appear all together on one page. BAMS is early for example. GFDL is late, meaning when you get it at 12Z for example, it's initialized position will be at 6Z. When comparing error data, it's not quite a fair comparison. NHC kind of covers that here:
http://www.nhc.noaa.gov/verification/verify6.shtml
The chart can kind of illustrate this too. For an early cycle model, the 0 hour error is comparing the current storm position to the initialization hour of that current model which was just released. The 0 hour for a late cycle model is actually compared to the storm position 6 hours ago. And everything is shifted by 6 hours for late cycle model data. Basically, you are not comparing like to like exactly. I don't really know how to explain that concept. There is no choice but to do it this way, but to be technically correct, I note it is not quite a fair comparison.

If none of this makes sense, that makes sense to me. I spent a month creating it and I still can't explain it well, lol.
68
In this thread:
Earl; Real-Time; Overnight - Tim_NC, 9/2/2010, 1:41 am
< Return to the front page of the: message board | monthly archive this page is in
Post A Reply
This thread has been archived and can no longer receive replies.