将 cumxxx (sum, min...) 应用于 DataFrame 中不同大小的 window

Question

我想对 DataFrame 中 不同大小 的 window 应用 cumxxx 操作。考虑以下输入：

import pandas as pd
from random import seed, randint
from collections import OrderedDict

p5h = pd.period_range(start='2020-02-01 00:00', end='2020-02-04 00:00', freq='5h', name='p5h')
p1h = pd.period_range(start='2020-02-01 00:00', end='2020-02-04 00:00', freq='1h', name='p1h')

seed(1)
values = [randint(0,10) for p in p1h]
df = pd.DataFrame({'Values' : values}, index=p1h)

p5h_st_as_series = p5h.start_time.to_series()

df['OpeneningPeriod'] = df.apply(
              lambda x: p5h.to_series().loc[p5h_st_as_series.index <=
                                            x.name.start_time].index[-1],
                                 axis=1)

结果

df.head(20)
                  Values   OpeneningPeriod
p1h                                       
2020-02-01 00:00       2  2020-02-01 00:00
2020-02-01 01:00       9  2020-02-01 00:00
2020-02-01 02:00       1  2020-02-01 00:00
2020-02-01 03:00       4  2020-02-01 00:00
2020-02-01 04:00       1  2020-02-01 00:00
2020-02-01 05:00       7  2020-02-01 05:00
2020-02-01 06:00       7  2020-02-01 05:00
2020-02-01 07:00       7  2020-02-01 05:00
2020-02-01 08:00      10  2020-02-01 05:00
2020-02-01 09:00       6  2020-02-01 05:00
2020-02-01 10:00       3  2020-02-01 10:00
2020-02-01 11:00       1  2020-02-01 10:00
2020-02-01 12:00       7  2020-02-01 10:00
2020-02-01 13:00       0  2020-02-01 10:00
2020-02-01 14:00       6  2020-02-01 10:00
2020-02-01 15:00       6  2020-02-01 15:00
2020-02-01 16:00       9  2020-02-01 15:00
2020-02-01 17:00       0  2020-02-01 15:00
2020-02-01 18:00       7  2020-02-01 15:00
2020-02-01 19:00       4  2020-02-01 15:00

此处，cumxxx 将应用于定义的 5 小时时段。它可以是不同的长度，因为 windows 可以是一天（有些带有夏令时），也可以是一个月（一个月中的小时数不是固定的）。

我要找的结果是：

df_result.head(11)
                  Values   OpeneningPeriod   Cumsum
p1h                                       
2020-02-01 00:00       2  2020-02-01 00:00        2  <- cumsum starts with a new period
2020-02-01 01:00       9  2020-02-01 00:00       11
2020-02-01 02:00       1  2020-02-01 00:00       12
2020-02-01 03:00       4  2020-02-01 00:00       16
2020-02-01 04:00       1  2020-02-01 00:00       17
2020-02-01 05:00       7  2020-02-01 05:00        7  <- cumsum starts with a new period
2020-02-01 06:00       7  2020-02-01 05:00       14
2020-02-01 07:00       7  2020-02-01 05:00       21
2020-02-01 08:00      10  2020-02-01 05:00       31
2020-02-01 09:00       6  2020-02-01 05:00       37
2020-02-01 10:00       3  2020-02-01 10:00        3  <- cumsum starts with a new period

与cummin & cummax也是同理。有人知道吗？

感谢您的帮助！最佳，

Answer 1

如果需要按 5H window 按 DatetimeIndex 分组，请使用 DataFrame.to_period 和 cumsum:

df['Cumsum'] = df.resample('5H')['Values'].cumsum()

或Grouper:

df['Cumsum'] = df.groupby(pd.Grouper(freq='5H'))['Values'].cumsum()

print (df.head(11))
                  Values   OpeneningPeriod  Cumsum
p1h                                               
2020-02-01 00:00       2  2020-02-01 00:00       2
2020-02-01 01:00       9  2020-02-01 00:00      11
2020-02-01 02:00       1  2020-02-01 00:00      12
2020-02-01 03:00       4  2020-02-01 00:00      16
2020-02-01 04:00       1  2020-02-01 00:00      17
2020-02-01 05:00       7  2020-02-01 05:00       7
2020-02-01 06:00       7  2020-02-01 05:00      14
2020-02-01 07:00       7  2020-02-01 05:00      21
2020-02-01 08:00      10  2020-02-01 05:00      31
2020-02-01 09:00       6  2020-02-01 05:00      37
2020-02-01 10:00       3  2020-02-01 10:00       3

Answer 2

groupby 应该是一个很好的起点：

df['Cumsum'] = df.groupby('OpeneningPeriod')['Values'].cumsum()

它给出：

                  Values  OpeneningPeriod  Cumsum
p1h                                              
2020-02-01 00:00       2 2020-02-01 00:00       2
2020-02-01 01:00       9 2020-02-01 00:00      11
2020-02-01 02:00       1 2020-02-01 00:00      12
2020-02-01 03:00       4 2020-02-01 00:00      16
2020-02-01 04:00       1 2020-02-01 00:00      17
2020-02-01 05:00       7 2020-02-01 05:00       7
2020-02-01 06:00       7 2020-02-01 05:00      14
2020-02-01 07:00       7 2020-02-01 05:00      21
2020-02-01 08:00      10 2020-02-01 05:00      31
2020-02-01 09:00       6 2020-02-01 05:00      37
2020-02-01 10:00       3 2020-02-01 10:00       3
2020-02-01 11:00       1 2020-02-01 10:00       4
2020-02-01 12:00       7 2020-02-01 10:00      11
2020-02-01 13:00       0 2020-02-01 10:00      11
2020-02-01 14:00       6 2020-02-01 10:00      17
2020-02-01 15:00       6 2020-02-01 15:00       6
...

将 cumxxx (sum, min...) 应用于 DataFrame 中不同大小的 window

Apply cumxxx (sum, min...) to a window of varying size in a DataFrame

python

pandas

cumsum

rolling-computation