Ignore 'NaN' values for calculation GEV

[Imran Hosen]

Dear All,

I am working with 3D precipitation daily data (time: 10950 lat: 172 lon: 221) and attempting to calculate Generalized Extreme Value (GEV) parameters using the scipy package. Below is the code I am using for this purpose:

# annual maximum precip
precip_annmax = precip.resample(time='Y').max()
precip_annmax = precip_annmax.chunk(dict(time = -1))
fit_GEV_numpy = lambda x: np.array(stats.genextreme.fit(x))
# fit the GEV distribution at each grid cell individually and in parallel
gev_params = xr.apply_ufunc(fit_GEV_numpy,
                            precip_annmax,
                            input_core_dims = [['time']],
                            output_core_dims  = [['param']],
                            output_sizes = {'param': 3},
                            dask = 'parallelized',
                            output_dtypes = [float],
                            vectorize = True)

Then, I'm trying to calculate the return period precipitation as follows:
precip_return50 = stats.genextreme.ppf(1 - (1/50),
                                       gev_params.isel(param = 0),
                                       gev_params.isel(param = 1),
                                       gev_params.isel(param = 2))

I am facing an issue with my dataset, which contains a significant number of NaN values. These NaN values seem to be causing problems, particularly when calculating the "return precipitation." I have already tried using the skipna=True option, but it hasn't resolved the issue.

Note: I also tested another dataset without NaN values and this code works fine. Only creates a problem when the dataset contains NaN values. 

Could you please suggest a way to properly ignore or handle NaN values during the calculation of the return levels?

Thank you for your time and suggestions.

Kind regards,

Imran