I have the aim to build a model to predict global horizontal irradiance (ghi) using satellite images and other features namely the day of the year and time of the day.
For extracting the satellite images features, I aim at using one of the available network in keras (e.g. Xception). Then I can concatenate the features output from the CNN to the other features (I already made the modifications to keras to handle those requirements.
Now, essentially, ghi is also a time-series problem so I assumed using time-series techniques could help. But I don't know where to look for those architectures (CNN + Time series + other features) so :
given my problem and desires, do you have ressources to help me tackle this issue?
one thing that is not yet clear to me is if I have don't freeze Xception internal layers, are the fiters parameters automatically learned by the model? I read here and there that filters are optimized also but I never found in keras a clear statement about that.