FELA · 1D FNO
Solar & Wind Power Forecasting
Setup
Probabilistic hour ahead renewable power forecasts from weather data
Pros
Runs offline in your browser · ~2 MB one time download
Usage
Pick a day, see the forecast vs power generation, no cloud (pun unintended!)
Weather... Forecasts?
A compact Fourier Neural Operator can map a short window of a numerical
weather forecast, each hour, to the distribution of renewable
power for that hour. Two tiny models: solar (137K parameters) and wind (312K parameters).
The shaded band is the model's uncertainty. The dots are what the power farm
actually generated.
A ... tiny forecast model?
These 'lil things are indeed the full size models, the browser runs them at
production fidelity, nothing is trimmed for the demo. The data is
real unseen test data from the GEFCom2014 competition set (June
2014 solar, December 2013 wind). So if your browser can predict real solar conditions
mostly accurately, what if wind turbines or solar cells could too? On site? No internet?
So what?
If the forecast can run on the site controller of a power farm, with realtime, up to date localized forecast data,
that same controller could significantly increase efficiency by knowing exactly when to charge and discharge storage systems.
Even in rural areas, if the network goes out, efficiency is maintained, correcting for error from supercomputer forecasts. It could also be used to stress test battery storage strategies
for upcoming extreme weather days to improve grid resilience! (Hey, maybe municipal power grids could stand to save money on Weather.com API bills!)
Full card on Hugging Face ↗