And now, here's Cli-Mate 9000 with the weather... Pattern-recognizing neural network tries its hand at forecasting Deep-learning software may help scientists predict extreme weather patterns more accurately than relying on today's weather prediction models alone. Simulations involving complex differential equations are run on supercomputers to predict the weather. The accuracy of forecasts using this approach have improved over time, …
Basically that's looking at synoptic situations and linking those to $TargetVariable. Interesting thing to do, and a logical thing to do, but not terribly surprising, because that's what companies do to "downscale" weather model output to the local scale - mostly (not only!) using classical statistics (though I'm not sure if they also do that for extremes as well, maybe?).
The clue is in the name, it's a model. Saying 'it’s also possible that the equations aren’t fully accurate" is not only hogwash but totally missing the point. The equations are absolutely not correct unless you are hacking into the reality mainframe and running on the matrix. And looking for patterns is, as said above, kind of obvious.
I for one welcome our robotic overlord Met presenters.
$ echo "Earlier on today, apparently, a woman rang the BBC and said she heard there was a hurricane on the way; well, if you're watching, don't worry, there isn't, but having said that, actually, the weather will become very windy, but most of the strong winds, incidentally, will be down over Spain and across into France."
$ out of cheese error
Well duh. We're talking thermodynamics, and there is not one lecture I've heard on the subject that doesn't start something like : "This subject is the most complex science in the history of Science".
I'm pretty sure the equations are not fully accurate, because they still can't model the actions of every single atom of matter in a given volume and, as long as we don't have the possibility of doing that, we do not have an accurate model.
On top of that, if you give the same calculation to a 16-bit processor, a 32-bit processor, a 64-bit processor and a 128-bit processor (supposing they exist or will soon), your model will soon provide 4 vastly different outcomes. So then you would think we just need to have 256-bit processors. You're not getting any closer to the "right" solution.
So we're just going to have to continue tweaking our existing models, knowing they are not, and cannot be, either perfect or reliable.
Edit : if you want a true explanation as to why we cannot possibly predict anything with perfect reliability, check this out.
Weather is something for which there is a wealth of accurate data, over time and geography, including extreme events, especially in first world countries. Why not train it on a subset of this data?
Until they get real, it's only useful to determine if I need a virtual umbrella.
As a weather weenie - their attempt to program in pattern recognition is exactly how human weather forecasts work. The guys and gals at the met offices look at all the main models, particularly those that are more accurate for their areas, mix that with their experience, and then create a forecast based upon a combination. Any service that relies strictly on one model is never as accurate as one that mixes data with the raw processing power and pattern recognition capacity of a person. And no method will ever be totally accurate because we will never know all the variables and how they interact every time. Chaos is never an easy equation to import into any forecasting system.
Weather prediction should be an extremely good match for machine learning approaches, in fact. What machine learning likes is lots and lots of training data, and weather has a huge amount of that. For instance: if you want to predict what will happen in 24 hours, well, you make the prediction, then you wait 24 hours, and you have the actual data of what really did happen which you can now use to train the system. And that goes on for ever: there is always more training data arriving because you always have data telling you what the conditions are now. And there's just vast amounts of data being puked down from all the sensors. So you get to train the system for ever.
Of course, the problem is that any trained NN-based model will probably be entirely opaque: you can't look inside it and find the bit where it's modelling clouds or something. But, for weather (not climate), that's fine: what matters is only how good the forecasts are, not that the numerical model that made them be comprehensible.
In a previous job I made just this suggestion, although I don't think they did anything about it. Presumably someone will, at which point the people running the big GCMs are going to look a bit silly, if it's not them who moved to machine learning.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
