pvfit5 — estimate SDM parameters at Standard Conditions from datasheet data using DEAP + pvlib

[valeriolobrano]

Hi everyone,

I'd like to share pvfit5, a Python package I developed for estimating the five parameters of the single-diode model at STC,
using only commercial datasheet values (Voc, Isc, Pmax, Vmp, Imp).

The approach uses a genetic algorithm (DEAP) to minimise the sum of relative errors on Voc, Isc, and Pmp. The I–V curve is
reconstructed via pvlib.pvsystem.calcparams_cec and the Lambert W single-diode solver.

I validated the tool against 13,000+ modules from the CEC database — the method and results are described in an Open Access
paper on Energy Reports: 10.1016/j.egyr.2026.109280

pip install pvfit5
pvfit5 --voc 36.3 --isc 8.19 --pmax 218.95 --vmp 29.0 --imp 7.55

It also works as a library:

from pvfit5.find_pv_parameters import fit_parameters, PVModuleData, STC, GAConfig
results, summary = fit_parameters(
PVModuleData(voc=36.3, isc=8.19, pmax=218.95, vmp=29.0, imp=7.55),
STC(), GAConfig()
)

Batch tools included: pvfit5-batch runs the algorithm on CEC modules, pvfit5-analysis and pvfit5-parametric produce
statistical analysis and per-technology breakdowns.

• GitHub: https://github.com/valeriolobrano/pvfit5
• PyPI: https://pypi.org/project/pvfit5/

Would love to hear feedback from the pvlib community. BSD-3-Clause licensed.

[cwhanse]

Hi Valerio, thanks for sharing information about this resource. To clarify, you write "the five parameters of the single-diode model" but a single-diode model (e.g., CEC) usually has more parameters than the five reference values for photocurrent, saturation current, shunt resistance, series resistance and diode quality factor. For the CEC model, there's also bandgap and its temperature coefficient, and a factor that is combined with the short-circuit temperature coefficient. It appears to me that users supply these other parameters and they are imposed on the estimation of the five reference parameters. Is that accurate?

[valeriolobrano]

Hi, great question, and you're absolutely right!
The name "five-parameter model" is really a historical convention in the accademic community. The five parameters being estimated are the classic ones: photocurrent (I_L), saturation current (I_o), series resistance (R_s), shunt resistance (R_sh), and the diode ideality factor (a). That's where the "five" comes from, and it's been called that way in the literature for decades.
Now, as you correctly noted, the full CEC single-diode model does involve additional quantities (bandgap energy (EgRef), its temperature coefficient (dEgdT), and the short-circuit current temperature coefficient (alpha_sc)). In pvfit5, these have sensible defaults (e.g., 1.121 eV and −0.000267 eV/K for crystalline silicon), but since the code is fully open source, users can adapt them freely to match their specific cell technology or experimental data.
This is actually a topic I've been working on for quite a while; my group published several papers on five-parameter estimation methods over the years:

• "An improved five-parameter model for photovoltaic modules"; Solar Energy Materials and Solar Cells, 2010
• "On the experimental validation of an improved five-parameter model for silicon photovoltaic modules"; Solar Energy Materials and Solar Cells, 2012
• "An efficient analytical approach for obtaining a five parameters model of photovoltaic modules using only reference data"; Applied Energy, 2013
  And pvfit5 itself is described in a recent open-access paper in Energy Reports (2026, DOI: 10.1016/j.egyr.2026.109280).
  The whole idea behind pvfit5 is to make this kind of estimation accessible and reproducible. Just plug in your datasheet values and let the genetic algorithm do the heavy lifting. But if you need to tweak the material parameters or the optimization strategy, everything is right there in the code.
  Happy to discuss further or hear suggestions from the pvlib community!

[markcampanelli]

I'd be interested in comparing to the spec-sheet fitting method I describe in "Comparison of calibrations of a single-diode model of photovoltaic module performance using performance matrix vs. specification datasheet" in Solar Energy, Volume 278, August 2024: https://www.sciencedirect.com/science/article/abs/pii/S0038092X24004079

It is an interesting (although perhaps sometimes ineffectual) endeavor to try to calibrate the SDM to a broad temperature and irradiance matrix from limited spec sheet information.

[cwhanse]

Hi Valerio, since the project references the CEC model, I am curious why the Adjust factor is omitted? Without Adjust, I think the underlying SDM is the De Soto model.

[valeriolobrano]

The literature contains many models that are supposed to outperform PVFIT5. The problem is that they are always described as having been tested and looking great, but there is no code available to actually test them in five minutes. The aim here has been to provide a practical procedure that anyone wishing to carry out tests. pvfit5 uses calcparams_cec from pvlib, and technically the CEC model differs from De Soto by the ddjust factor which is a percentage correction to alpha_sc for translating parameters to non-STC conditions.
However, pvfit5's scope is specifically the estimation of the five reference parameters from STC datasheet data. So for the purpose of finding the five reference parameters, the distinction between CEC and De Soto there is not. Once you have the fitted parameters (I_L_ref, I_o_ref, R_s, R_sh_ref, a_ref), you're free to use them with either model in pvlib with whatever adjust value is appropriate for your module!

[markcampanelli]

Hopefully this Getting Started code for PVfit hasn't gotten stale:

https://github.com/markcampanelli/pvfit/blob/main/pvfit/modeling/dc/single_diode/model/demos/getting_started.py

This code compares fitting PVfit's "simple" SDM using IEC 61853-1 matrix data to using only spec-sheet data for the same module.

Happy to try to improve usability where needed.

[adriesse]

Hi Valerio,

Thanks for publishing and starting the discussion!

Maybe the answer to my question is in your paper (which I intend to read later) but I was a bit puzzled by your objective function:

E = |Voc_sim − Voc*|/Voc* + |Isc_sim − Isc*|/Isc* + |Pmp_sim − Pmp*|/Pmp*

It seems you are not trying to aim for the correct Vmp and Imp, but only their product. Why is that? Hopefully they match of course, but I don't think they always will.

On your graphical output it might be a good idea to show both points, datasheet and fitted. That way the user can be alerted to any discrepancies.

[valeriolobrano]

Hi. The objective function shown in the paper is a simplified description of the earlier version. The current implementation in pvfit5 actually uses a weighted multi-objective that includes Vmp and Imp separately: E = 5·e_Voc + e_Isc + e_Vmp + 15·e_Imp + 0.3·stationarity_residual So Vmp and Imp are individually targeted — not just their product. The heavy weight on Imp (15×) reflects the fact that Imp is typically the hardest parameter to match precisely with the single-diode model. On top of that, there's a stationarity constraint that checks dP/dV ≈ 0 at the datasheet Vmp. This is a physically meaningful condition: at the true maximum power point, the power derivative must vanish. It acts as an additional anchor ensuring the fitted MPP lands at the right voltage. The output summary already reports Vmp(sim) vs Vmp* and Imp(sim) vs Imp* individually, so the user can immediately spot any discrepancy. Regarding the plot — that's a really good suggestion! Icould add the fitted (Vmp_sim, Imp_sim) point alongside the datasheet point on Graph 2, so any mismatch is immediately visible.