issues
search
openclimatefix
/
power_perceiver
Machine learning experiments using the Perceiver IO model to forecast the electricity system (starting with solar)
MIT License
7
stars
1
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
Don't include historical pv "prediction" in the objective function
#99
JackKelly
closed
2 years ago
0
Try using an RNN for PV (followed by transformer encoder)
#98
JackKelly
closed
2 years ago
0
Give the time_transformer 3 different inputs for PV.
#97
JackKelly
opened
2 years ago
0
Add back in the GSP timestep to the GSP query generator for satellite transformer
#96
JackKelly
closed
2 years ago
0
"Aux" output for 5-minutely GSP predictions from `SatelliteTransformer`
#95
JackKelly
opened
2 years ago
0
Try having an "aux output" directly from the `SatelliteTransformer` into a FF net, to directly predict PV from the `SatelliteTransformer`.
#94
JackKelly
closed
2 years ago
0
Give SatelliteTransformer and time_transformer the t0 time encoding
#93
JackKelly
closed
2 years ago
0
Train just the u-net for a few epochs. Then train the whole model.
#92
JackKelly
opened
2 years ago
0
Improve model performance
#91
JackKelly
opened
2 years ago
0
Enlarge topography data so we never get NaNs over land
#90
JackKelly
opened
2 years ago
0
Copy raw PV and GSP data (and any aux files needed) to GCP VM
#89
JackKelly
closed
2 years ago
0
Get model working with new data
#88
JackKelly
closed
2 years ago
0
Why is the topo data blank?
#87
JackKelly
closed
2 years ago
1
If there are no PV systems, then fill with NaNs, and don't check for NaNs!
#86
JackKelly
closed
2 years ago
0
Collate full batch within `RawDataset`. Pass batch to `np_batch_processor`
#85
JackKelly
closed
2 years ago
0
Why does `assert (datetime_index == datetime_index.round("5T")).all()` pass, even though satellite dates aren't rounded to 5 mins?
#84
JackKelly
closed
2 years ago
1
Create notebook to plot data from new `RawDataset`
#83
JackKelly
closed
2 years ago
0
When giving the model days of NWPs, mask nighttime on the fly, and don't load nighttime
#82
JackKelly
opened
2 years ago
0
Compute Union of periods before loading into ram
#81
JackKelly
opened
2 years ago
0
Need to think about how multiple days of NWPs fit with the idea of contiguous periods
#80
JackKelly
opened
2 years ago
1
Maybe save prepared PV data to speed up tests and loading
#79
JackKelly
opened
2 years ago
0
Only have one of each RawDataSource in RawDataset
#78
JackKelly
opened
2 years ago
0
Tell the model which satellite pixels are in the target GSP
#77
JackKelly
opened
2 years ago
0
Try using `gsp_id_to_shape.geometry.envelope.centroid`
#76
JackKelly
closed
2 years ago
3
Give the model the installed GSP PV capacity
#75
JackKelly
opened
2 years ago
0
Give the model the "raw" PV data
#74
JackKelly
opened
2 years ago
0
Don't repeat PV systems. Instead, pad with NaNs and mask the loss
#73
JackKelly
opened
2 years ago
0
Check that all the processing code can cope with NaN empty examples
#72
JackKelly
closed
2 years ago
0
Mask the loss to include the correct examples per batch per data source combo.
#71
JackKelly
closed
2 years ago
0
Check that fourier encoding is done on a per-example basis
#70
JackKelly
closed
2 years ago
1
implement code to get an empty example
#69
JackKelly
closed
2 years ago
0
Implement `RawGSPDataSource`
#68
JackKelly
closed
2 years ago
0
Implement `RawPVDataSource`
#67
JackKelly
closed
2 years ago
0
get `satellite_zarr_dataset` working with the new dataset before working on GSP or PV.
#66
JackKelly
closed
2 years ago
0
train directly from the raw GSP and PV data
#65
JackKelly
closed
2 years ago
0
BUG: Why do the training plots have different numbers of PV systems in the actual vs predicted?!
#64
JackKelly
closed
2 years ago
1
Combine UNet with Power Perceiver
#63
JackKelly
closed
2 years ago
0
Run U-Net in "MetNet mode": Predict one timestep at a time. Tell the model which timestep by using an input channel. Output a probability distribution. Only predict for the central crop.
#62
JackKelly
opened
2 years ago
0
Ideas for after productionisation
#61
JackKelly
opened
2 years ago
0
Use Sun's angle to decide which times are "daylight"
#60
JackKelly
closed
2 years ago
2
Train from Zarr
#59
JackKelly
closed
2 years ago
0
Implement ClimateHack.AI model
#58
JackKelly
closed
2 years ago
0
Only keep the best model version on wandb
#57
JackKelly
opened
2 years ago
1
Use a subset of timesteps during training
#56
JackKelly
closed
2 years ago
0
Implement `AlignGSPTo5Min`
#55
JackKelly
closed
2 years ago
0
Pass through both models without any `for` loops
#54
JackKelly
closed
2 years ago
0
wandb tweaks
#53
JackKelly
closed
2 years ago
0
Plug "power-from-imagery Transformer Encoder"'s predictions into a second Transformer Encoder which looks across time.
#52
JackKelly
closed
2 years ago
1
Include solar azimuth and elevation in the GSP query
#51
JackKelly
opened
2 years ago
1
Transformer Encoder which infers PV power for GSP regions & individual PV systems from several satellite images (but doesn't predict the future)
#50
JackKelly
closed
2 years ago
0
Previous
Next