預測 Forecast vs. actual (synthetic test set, 10-min steps)
actual model (computed live in JS) persistence
風 Try it: next-step forecast
Forecast (next 10 min)–
Persistence–
The sliders build a 24-step history ending at the chosen speed; the power columns follow a generic cubic power curve.
衡 Test-set metrics (synthetic)
| Model | Persistence |
|---|
Persistence predicts "next = current". At 10 minutes ahead it is already very accurate, so it is the honest yardstick.
道 How it works
Dilated causal convolutions (a TCN) and self-attention produce P10/P50/P90 quantile forecasts; the band can drive an advisory, human-approved reserve suggestion. The model predicts the 10-minute change in wind speed and adds it to the current reading.
Weights were trained in PyTorch, exported to JSON and re-implemented in plain JavaScript. A test checks that the JS and Python predictions match within 1e-4.