• Welcome to TalkWeather!
    We see you lurking around TalkWeather! Take the extra step and join us today to view attachments, see less ads and maybe even join the discussion.
    CLICK TO JOIN TALKWEATHER
  • Current Tropical Systems
    Erin

2025 Atlantic Hurricane Season

wx_guy said:
This is an unprecedented level of consistency for an AI model that far out -- this is, once again, a huge test for the AI models long range. The AI models are trained on historical patterns, so it's curious to see the very strong consistency. I am eagerly waiting for other models to "fall into line" on this, but it is a noteworthy amount of steady forecasting. I know the Euro AI and DeepMind AI are new, but do we have any verifications on their errors yet?
Click to expand...
EURO and GFS see this as an east coast threat too, so it really concerns me.
 
KakashiHatake2000 said:
wait hold on atlantic that is a 360 hour run yeah its from the euro ai but its still too far out since it is 360 hours out things can change in 360 hours
Click to expand...
I mean obviously and I did emphasize that in the global thread by saying “hypothetical”
 
IMG_3224.png

im not sure if this would be related to tropics or not
 
KakashiHatake2000 said:
IMG_3224.png

im not sure if this would be related to tropics or not
Click to expand...
That’s alarming. Back in January and February I was considering 2005 as one of my analogs, but I dropped it for 2008 and 2017.
 
That 2005 analog did come thru my mind after seeing the 12z Euro Operational run.
 
Here are interesting recent articles discussing the accuracy of AI models in tropical cyclones.




www.nature.com

Benchmark dataset and deep learning method for global tropical cyclone forecasting - Nature Communications

The authors present TropiCycloneNet, a framework that combines a multimodal tropical cyclone dataset spanning 70 years and a machine learning forecast model. The approach improves forecast skill of tropical cyclone track and intensity compared to other methods, advancing data-driven weather...
www.nature.com www.nature.com


It seems that the AI models are a significant improvement on track forecasting over the traditional NWP deterministic models. This is great, of course, but the "big unknown" and the "holy grail" of TC forecasting -- rapid intensification prediction -- still eludes us. At least for now. The AI models so far have seen to significantly *under forecast* TC intensity and are much less accurate at intensity forecasting than NWP deterministic models so far. This makes sense, structurally, as the AI models (deep learning neural networks) have a very hard time extrapolating beyond the observed patterns in their datasets. RI might represent 3%-7% of actual TC activity, so that rarity makes it difficult to explicitly forecast in the statistical AI modeling that occurs. I think the push over the next few years will be to incorporate rare-event modeling into them to try and capture these rare events, but it's a very difficult problem to accurately fix from a statistical standpoint. It might well be that advances in NWP deterministic models and tools like SHIPS advancing in the future becomes the "intensity" forecast component while AI modeling focuses on the track. In any case, it'll be an interesting next few. years as we meld together traditional and new approaches in the field.
 
wx_guy said:
Here are interesting recent articles discussing the accuracy of AI models in tropical cyclones.




www.nature.com

Benchmark dataset and deep learning method for global tropical cyclone forecasting - Nature Communications

The authors present TropiCycloneNet, a framework that combines a multimodal tropical cyclone dataset spanning 70 years and a machine learning forecast model. The approach improves forecast skill of tropical cyclone track and intensity compared to other methods, advancing data-driven weather...
www.nature.com www.nature.com


It seems that the AI models are a significant improvement on track forecasting over the traditional NWP deterministic models. This is great, of course, but the "big unknown" and the "holy grail" of TC forecasting -- rapid intensification prediction -- still eludes us. At least for now. The AI models so far have seen to significantly *under forecast* TC intensity and are much less accurate at intensity forecasting than NWP deterministic models so far. This makes sense, structurally, as the AI models (deep learning neural networks) have a very hard time extrapolating beyond the observed patterns in their datasets. RI might represent 3%-7% of actual TC activity, so that rarity makes it difficult to explicitly forecast in the statistical AI modeling that occurs. I think the push over the next few years will be to incorporate rare-event modeling into them to try and capture these rare events, but it's a very difficult problem to accurately fix from a statistical standpoint. It might well be that advances in NWP deterministic models and tools like SHIPS advancing in the future becomes the "intensity" forecast component while AI modeling focuses on the track. In any case, it'll be an interesting next few. years as we meld together traditional and new approaches in the field.
Click to expand...
Like all things, a tool in the toolbox, but not a farmer's almanac. Fascinated to see where these go in terms of aiding forecasting - an instance where machine learning can be used to actually save lives and improve community outcomes.
 
You must log in or register to reply here.
Back
Top