How to stack timeseries into period with overlap

[saschahofmann]

I would like to split/stack a timeseries into various periods e.g. imagine a 100 year timeseries. That I would like to split into periods of 20 years. I can do this e.g. like this

import xarray as xr 
import numpy as np
from datetime import timedelta

def stack_periods(data, period_length=20):
    periods = []
    times = xr.date_range(
        start="19000101T1200",
        periods=period_length * 360,
        freq="D",
        calendar="360_day",
    )
    for period, da in data.resample(time="20YS"):

        center_period = period + timedelta(days=period_length // 2 * 360)
        periods.append(da.expand_dims(period=[center_period]).assign_coords(time=times))

    return xr.concat(periods, dim="period")

time = xr.date_range('2000-01-01','2099-12-30', freq='D', calendar='360_day')
data = xr.DataArray(np.random.randn(time.size), coords=dict(time=('time', time)))

stack_periods(data)

As you can see, I am creating an artificial time series since the periods dont really have the same timestamps.

Now what I couldnt figure out is, how to generate periods that overlap - e.g. 20 year periods but for every 10 years. I thought .rolling+.construct should somehow do the trick but I am missing an option to center the rolling groups on the left.

data.rolling(
        time=7200, center=True
    ).construct(window_dim="time_period", stride=7200)

The data is shifted I think by 10 years that's why there are NaNs in the beginning.

I guess I could somehow shift or add data before doing the rolling, to make that work but maybe there is a more straightforward way to achieve this?

[dcherian]

So you want center=True but a different coordinate? If so, you'll have to do adjust the coordinates manually.

[saschahofmann]

What I would want is the rolling period to start offsetted e.g. 10 years, with center = False the first 10 years are null. If I set it to False the first 20 years are null.

Another option would be to center on the right or something like this?

[saschahofmann]

Basically, I would like a way to create the same output as this would produce with the flexibility to change the stride (in this case 10 years)

xr.concat(
    [stack_periods(data), stack_periods(data.isel(time=slice(7200, 360 * 100 - 7200)))],
    dim="period",
).sortby("period")

is there any way to do that using rolling?