Add flag for when pv generation is NaN

[zakwatts]

Currently in the sites pipeline, when there is no generation data for sites, we infill any NaNs with 0's.

This as @JackKelly has stated is quite "mean" to the model, expecting it know when we don't have site generation data but are passing it 0 values.

One solution to this could be to change the value from 0, to a very specific number that the model could learn that when site generation is that value that there is no real generation.

Another solution which I propose in this issue (thank you @Sukh-P) for the idea, is to pass a flag to the model for each site generation for when it is available. This could be an extra datavariable which is 1, for when site generaiton exists and 0 for when we have replaced NaNs in the dataset for 0. This hopefully allow the model to learn when there is No PV generation better.

Open to thoughts on other solutions and how this could be implemented.

[zakwatts]

Currently we fill NaNs with 0 here:

ocf-data-sampler/ocf_data_sampler/torch_datasets/utils/merge_and_fill_utils.py

Line 15 in 2ea62dc

def fill_nans_in_arrays(sample: dict) -> dict:

Currently the function np.nan_to_num is used which infills NaNs with 0: https://docs.scipy.org/doc//numpy-1.13.0/reference/generated/numpy.nan_to_num.html

[Sukh-P]

Thanks for opening this Zak! Just an FYI this is partially done by this PR here #354 in terms of making the fill value configurable and can be set to a different value apart from 0, other parts of adding a binary flag indicating dropped out values is still to do

[zakwatts]

Thanks @Sukh-P, I do some experimentation using -1 as the dropout value, as this might allow the model to capture the relationship