Overflow in clearsky.haurwitz when apparent_zenith is almost 90

[cwhanse]

Haurwitz model includes this term:

np.exp(-0.059/cos_zenith)

We currently don't intercept and handle numerical overflow when cos_zenith is very close to 0.

[wholmgren]

Cliff, I don't fully understand what you mean by numerical overflow in the python/numpy context since you usually will get the largest number your platform supports, a nan, or an inf. In any case, it might be platform and numpy version dependent. On my Python 3.5, 64-bit mac and my Python 2.7, 64-bit linux, both running numpy 1.10.4...

In [2]: np.cos(np.deg2rad(89.99999999999999999999999999999999999))
Out[2]: 6.123233995736766e-17

In [3]: 1/np.cos(np.deg2rad(89.99999999999999999999999999999999999))
Out[3]: 16331239353195370.0

In [4]: np.exp(-0.059/np.cos(np.deg2rad(89.99999999999999999999999999999999999)))
Out[4]: 0.0

Same results with 90 and 90.0, and the behavior from 89.0 - 89.9 also seems ok. Can you copy/paste a few lines of input and output?

[cwhanse]

Hi Will,

Perhaps it’s a runtime setting on my end. I get warnings, not errors, when the zenith is 90 (or very close to it):

C:\Users\cwhanse\AppData\Local\Continuum\Anaconda3\lib\site-packages\pandas\computation\expressions.py:62: RuntimeWarning: invalid value encountered in power
return op(a, b)
C:\Users\cwhanse\AppData\Local\Continuum\Anaconda3\lib\site-packages\pvlib\clearsky.py:314: RuntimeWarning: overflow encountered in exp
clearsky_GHI = 1098.0 * cos_zenith * np.exp(-0.059/cos_zenith)

The code executes to end. Matlab handles this case with the ‘Inf’ value, which made it transparent to us.

I mostly wanted to learn how to post an issue, you can close it if it’s not really an issue.

Cliff

[wholmgren]

Are you using Spyder on Windows? I wonder if it's related to pandas-dev/pandas#9950 and links therein. You might be able to eliminate the warnings by running np.errstate(over='ignore') or, more aggressively, np.errstate(all='ignore')

[wholmgren]

Actually, I think np.seterr might be the right one to use.