I would really like the codes...

[tcapelle]

I am trying to reproduce this papers results on with little success.
Could you provide codes and/or datasets?
SIncerely,
Thomas

[talha111]

You can scrap the data from

http://skycam.mmto.arizona.edu/
https://midcdmz.nrel.gov/apps/skygallery.pl

Unfortunately we are not able to get clearance to share the code.

[tcapelle]

Thanks for the response,
Coudl you provide more details on training?
I am unable to get good results with the LSTM layers setup described on the paper.

• Did you use stateful LSTMS?
• Did you organized the train dataset on some particular order to feed the model, like it is done with language model, shuffling with order, chunk-ifying the data, etc...
• How many LSTM layers are on each?
• How many forward steps you predicted at training time. Did you use some weighted MSE to account for the forecasting importance?
• Why you did not use a more traidtional Resnet encoder, i find better performance with a more standard model...
• Did you benchmarked against the ConvLSTM layers (available in Keras).

Sorry so many questions, but I have not been able to get good results with this model. But I have the intuition that it should perform better.

Sincerely,
Thomas