Hey, if you’re trying to catch those quick wind shifts, I’d lean toward a model that can track time really well. Long‑short‑term memory networks or gated recurrent units are classic for sequences, but lately people use Temporal Convolutional Networks or even Transformer‑style models that can look ahead and behind in the data. For microbursts, where the signal can be a few seconds, a lightweight GRU or a Temporal ConvNet with a few layers might give you the best trade‑off between speed and accuracy. Just make sure you feed it a good spread of past and future points so the network can learn the real pattern, not just noise.

A 30‑ to 60‑second window usually works well. If your data are in one‑second intervals that’s 30–60 timesteps, long enough to capture the pre‑burst buildup but short enough to keep the model fast. If you see the burst signature is very brief, you could trim it to 20–30 seconds and let the GRU learn the sharp spike. Good luck!

Sounds like a solid plan. Watching the loss curves will definitely catch any over‑fitting early. Let me know if the 60‑second window pulls in more useful context or if the shorter window keeps things tighter. Good luck!

You’re doing great—keep listening to the data and let it guide you. Good luck!