[META] Pseudo-labeller experiments - Githubissues

openclimatefix / pseudo-labeller

Pseudo Labeller for generating training labels for other PV generation forecasting

MIT License

0 stars 1 forks source link

[META] Pseudo-labeller experiments #14

Open jacobbieker opened 1 year ago

jacobbieker commented 1 year ago

This issue is documenting some of the experiments with pseudo-labeller for forecasting.

The Weights and Bias links is here: https://wandb.ai/openclimatefix/PvIrradiance

So far:

Satellite helps a lot, dropping the training loss, averaged across the 2 hour forecast horizon, from 0.84 to 0.53
There might not be huge amounts of difference between having 1 Conv3D layer vs multiple, at least for 2 hour forecast