Windninja fix

[scotthavens]

Wind Ninja fix

The fix for the wind ninja bug is a little more elegant and has moved the smrf.distribute.wind to a package where you can select what wind model to use. A config option wind_model has the following values:

• winstral the maximum upwind slope distribution
• wind_ninja to load wind ninja outputs
• interp standard interpolation

The wind ninja performs more robust linear interpolation along the edges and will raise an error if there are still NaN values in the image. To test this, I've added the Lakes domain with a HRRR simulation that mirrors what we do in real time, something that we were not testing before.

Other updates

• Fix for issue #157, got lazy and just rolled it in
• Added tests for the recipes in the wind section to ensure that the output config file is how it should look like without additional item
• some of the previous changes required pandas>1.0 so updated the requirements
• Also fixed the weather_forecast_retrieval to be the newer version
• Added a notebooks folder to house jupyter notebooks that can be used to look at SMRF outputs, currently just looking at the gold files
• netCDF4>1.4.0 will return masked numpy arrays, apply a function that will convert to normal numpy array if no fill values are present
• In the tests there was a redundant gridded data test, it was the same in test_load_data.py and test_gridded_data.py

Changes in the gold files

The RME gold and gold_hrrr changed.

Above are the changes to the RME gold files. Almost no changes expect for some small changes in net_solar which was expected from the slope fix.

Above are the changes to the RME gold_hrrr files. Most of the changes appear to be float tolerances which was shown to be an issues with how xarray in weather_forecast_retrieval reads grib files. The net_solar shows the largest change from the slope fix but very minimal.

[jomey]

Really like the break up to a several wind classes.

Have you done any performance testing on this? I would be curious to see how it performs on big datasets as I see a lot of numpy operations that could be vectorized and probably get a big boost this way.

[scotthavens]

The breakup of the wind classes was necessary, it would have been to large of a class without it. I'm hoping that this will serve as a template for some of the other classes we have to start compartmentalizing the code a bit more. Plus this makes it a lot easier to unit test without having to run all of SMRF.

There was a point I was testing the WindstralWindModel on a GPU but didn't make a lot of progress. I bet with numba we could see an increase in performance.