search

openclimatefix / PVNet

PVnet main repo
MIT License
17 stars 3 forks source link

Upgrades to PVNet #2

Closed peterdudfield closed 1 year ago

peterdudfield commented 1 year ago
peterdudfield commented 1 year ago

@dantravers @JackKelly @jacobbieker I thought I would write down the improves that could be (easily) made to PVNet, I think this lead on from our discussion yesterday

dantravers commented 1 year ago

Thanks @peterdudfield . Here's a screenshot I took yesterday whcih shows the 4 hour forecast. I've been noticing it seems to significantly underestimate the national outturn up to ~10.30, and then over-estimate after that. So the forecasting being generated at 6.30am seems to get new information. Not sure where from.
image

dantravers commented 1 year ago

Comments on the proposed changes: Add error bars on results - multiple Gaussian distribution DT: Would be good to get probabilistic, but I'm aware that it will need UI and API changes to show the impact, so not sure if I'd prioritise this. Train with more NWP data (currently only 2 hours) DT: Is that 2 hours into the future? If so - definitely sounds like would be useful. Train with more data DT: not sure what data, but always sounds good! Check night time data Add more NWP channels, currently only dswrf DT: From my research - temperature is the next most important variable. Could help, but I suspect isn't the main driver of changes.

peterdudfield commented 1 year ago

https://app.neptune.ai/o/OpenClimateFix/org/predict-pv-yield/e/PRED-1231/charts This is using 4 NWP variables and using 4 hours of NWP data

peterdudfield commented 1 year ago

upgrade forecast with NWP data -https://github.com/openclimatefix/nowcasting_forecast/issues/176

dfulu commented 1 year ago

The library has moved on quite a lot since this, so I'm going to close this as stale. However, the current model addresses most of these:

Add error bars on results - multiple Gaussian distribution

Currently underway via the quantile regression support #44

Train with more NWP data (currently only 2 hours)

Implemented. Using from -2 up to +8 hours for forecast

Train with more data

Currently training on about 1,600,000 samples

Check night time data

In production this is done inside the app

Add more NWP channels, currently only dswrf

Currently we only use dswrf and t, so further exploration would be a good idea