Wrong implementation of Louche decomposition?

[ramaroesilva]

Describe the bug
According to pvlib's source code, the model explores an empirical regression between DNI/I_0 and GHI/I_0 (i.e., Kt).

However, the original paper states

For direct normal radiation (noted I), we used as a reference, a calculated direct normal irradiation for cloudless skies (noted I_M), based on the model elaborated by A. Louche et al.[13].

Meaning that they use normalize DNI by a clear-sky version rather than the extraterrestrial.

I've looked at the [13] paper and it was not very clear, as it is mainly oriented to estimating turbidity. The only more similar thing I found was

There is also the fact that the current implementation does not saturate to 0 for Kt<0.17 as stated in the paper. As is, if Kt=0, it leads to one of the components being negative.

Lastly, maybe it could also be mentioned in the doc - assuming that the correct implementation is viable within pvlib -that this model was originally made for daily averages.

[cwhanse]

I think I disagree here. The polynomial implemented in pvlib is not the quadratic you display, but the 5th degree polynomial on the following page. The quadratic is for daily ratios (DNI to clear-sky DNI); the 5th degree polynomial is for hourly ratios. The text preceding the 5th degree polynomial states

"For hourly values, we cannot use I_h/I_M,h because the model (my comment: here, model refers to the daily quadratic I think) is based on monthly mean values...So, we used, as a reference, the hourly extraterrestrial normal irradiance".

I can understand the confusion in that both polynomials use I_M but I think the meaning is different: in the quadratic, I_M is from the clear-sky model described in [13], while in the 5th degree polynomial it is the extraterrestrial reference.

This function was ported from PVLib for Matlab, and the comments in Matlab don't address this question.

[ramaroesilva]

@cwhanse thank you for the careful verification and sorry for wasting your time, you're absolutely right. I was clearly tired yesterday, I didn't even notice that the expressions were different, just focused on the variables it used...

I overreacted to getting a negative dhi when ghi=0 in a code where I had yet to include some qc indexing like solar elevation>min_thresh or ghi>min_thresh.

Louche's hourly expression has a bias term of +0.02 and pvlib's kt calculation is never nan even when solar zenith ≥ 90º*. So I was getting kt=0 during nighttime for which Louche produces a positive dni and, as a consequence, a negative dhi (for consistency with ghi=0). For the sake of sharing, this happened for GHI up to ~0.5 W.m-2.

*I0h = extra_radiation * np.maximum(cos_zenith, min_cos_zenith), which means an always positive cos_zenith

Apologies once again and thank you for the always insightful discussions!

[echedey-ls]

IMHO, still worth some improvements in the code, comments specifically, to avoid this kind of confusion and ease future reviews like @ramaroesilva did.

Edit to add: well, a notes section or an admonition to explain non-positive outputs may as well be added.

[ramaroesilva]

You mean with the docstrings or with the code itself?

Although a QC filtering would have avoided all this, maybe it could make sense to block negative values stating that this goes beyond the original paper.

[echedey-ls]

Both. Inline comments in the code (like, # Eq. (x) from [1], and the Notes section or admonition anywhere in the docstring to state the negative values that can be yielded.

[cwhanse]

handling negative dhi:

this goes beyond the original paper.

I agree here. We could put a floor on the output, preserving nan when nan are in the input.

As far as traceability back to the reference, I'd vote for inline comments over a Notes section.

[echedey-ls]

Reopening so it's not forgotten.

[ramaroesilva]

@echedey-ls, was thinking of keeping this issue closed and opening a new, better framed issue since the starting point for this discussion came from me misreading the original paper (while mentioning it started from exchanges here).

Is that ok?

[echedey-ls]

I don't see much difference, please do whatever is more comfortable to you. My unique hesitation is that AI bots scrap for new issues and if I review one more microslop hallucination I drop my career in software and retire to a small village in the countryside, buy a tractor and farm for the rest of my life (jk I don't have the money for that).