Data driven model

[jacktang]

Hello, now we setup pvlib simulation by determined physical parameters, and i wonder is there any alternative to train the pvlib model by existing generation data and weather data?

[echedey-ls]

pvlib does not offer that, as it's only an aggregation of models and wrappers to ease their use. I recommend you use your historical data to do that manually, choosing the models and values appropriate for your plant. And by using the same ones in the future, you can have an insight on degradation.

There are other tools out there, that may help you when only the power output timeseries is known. E.g.: https://solar-data-tools.readthedocs.io/en/stable/

[cwhanse]

Assuming you have measured weather and power data, and these data are aligned in time, I think you could estimate parameters for a basic weather-to-power model such as pvlib.pvsystem.pvwatts_dc.

If you plot the power vs. time, you can tell from the shape if the system is fixed tilt or tracking. Fixed tilt has a sine wave profile, tracking is more "square". Try it with a simulated system and clear-sky irradiance and you'll see what I mean. You will also see if the system is clipping at high power.

For a fixed tilt system, the model process would be:

1. Translate measured irradiance to plane of array. It sounds like you don't know the array configuration, so there are two unknowns: surface_tilt and surface_azimuth.
2. Convert air temperature to module temperature. You can use any of the temperature models with default parameters for a fixed-tilt, free-standing system.
3. Translate from plane of array irradiance and module temperature to DC power using pvwatts_dc. There are two more unknowns: DC power at standard conditions (1000 W/m2, 25C) pdc0, and the temperature coefficient of power gamma_pdc.
4. I wouldn't worry about an inverter model, only the clipped power level you observe in the power data.
5. Set up an optimization where you estimate the four unknowns (tilt, azimuth, pdc0, gamma_pdc) from the input data (GHI, air temperature, AC power).

This process should produce a model that gets within 10%, and perhaps closer, to true power. There are ways to refine this estimation process, e.g., use something like pvlib.clearsky.detect_clearsky on either the GHI or power data to isolate periods of clear sky and fit the model to these data. Or plot power vs. hour of day and use that to estimate the azimuth.

Let us know what you learn as you pursue this.