将不规则时间戳的测量值转换为等间距的时间加权平均值

2022-01-11 00:00:00 python pandas time-series

问题描述

我有一系列带有时间戳且间隔不规则的测量值.这些系列中的值始终代表测量值的变化——即没有变化就没有新值.此类系列的一个简单示例是:

I have series of measurements which are time stamped and irregularly spaced. Values in these series always represent changes of the measurement -- i.e. without a change no new value. A simple example of such a series would be:

23:00:00.100     10
23:00:01.200      8
23:00:01.600      0
23:00:06.300      4

我想要达到的是一系列等距的时间加权平均值.对于给定的示例,我可能会针对基于秒的频率,因此结果如下:

What I want to reach is an equally spaced series of time-weighted averages. For the given example I might aim at a frequency based on seconds and hence a result like the following:

23:00:01     NaN ( the first 100ms are missing )
23:00:02     5.2 ( 10*0.2 + 8*0.4 + 0*0.4 )
23:00:03       0
23:00:04       0
23:00:05       0
23:00:06     2.8 ( 0*0.3 + 4*0.7 )

我正在寻找解决该问题的 Python 库.对我来说,这似乎是一个标准问题,但目前我在 pandas 等标准库中找不到这样的功能.

I am searching for a Python library solving that problem. For me, this seems to be a standard problem, but I couldn't find such a functionality so far in standard libraries like pandas.

算法需要考虑两件事:

  • 时间加权平均
  • 在形成平均值时考虑当前间隔之前的值(甚至可能在领先者之前)
data.resample('S', fill_method='pad')          # forming a series of seconds

做部分工作.为聚合提供用户定义的函数将允许 形成时间加权平均值,但是因为忽略了区间的开始,所以这个平均值也是不正确的.更糟糕的是:系列中的孔被平均值填充,在上面的示例中导致第 3、4 和 5 秒的值不为零.

does parts of the work. Providing a user-defined function for aggregation will allow to form time-weighted averages, but because the beginning of the interval is ignored, this average will be incorrect too. Even worse: the holes in the series are filled with the average values, leading in the example from above to the values of seconds 3, 4 and 5 to be non zero.

data = data.resample('L', fill_method='pad')   # forming a series of milliseconds
data.resample('S')

以一定的准确性完成这个技巧,但是 - 取决于准确性 - 非常昂贵.就我而言,太贵了.

does the trick with a certain accurateness, but is -- depending on the accurateness -- very expensive. In my case, too expensive.

import pandas as pa
import numpy as np
from datetime import datetime
from datetime import timedelta

time_stamps=[datetime(2013,04,11,23,00,00,100000), 
             datetime(2013,04,11,23,00,1,200000),
             datetime(2013,04,11,23,00,1,600000),
             datetime(2013,04,11,23,00,6,300000)]
values = [10, 8, 0, 4]
raw = pa.TimeSeries(index=time_stamps, data=values)

def round_down_to_second(dt):
    return datetime(year=dt.year, month=dt.month, day=dt.day, 
                    hour=dt.hour, minute=dt.minute, second=dt.second)

def round_up_to_second(dt):
    return round_down_to_second(dt) + timedelta(seconds=1)

def time_weighted_average(data):
    end = pa.DatetimeIndex([round_up_to_second(data.index[-1])])
    return np.average(data, weights=np.diff(data.index.append(end).asi8))

start = round_down_to_second(time_stamps[0])
end = round_down_to_second(time_stamps[-1])
range = pa.date_range(start, end, freq='S')
data = raw.reindex(raw.index + range)
data = data.ffill()

data = data.resample('S', how=time_weighted_average)


解决方案

您可以使用 traces 来做到这一点.

You can do this with traces.

from datetime import datetime
import traces

ts = traces.TimeSeries(data=[
    (datetime(2016, 9, 27, 23, 0, 0, 100000), 10),
    (datetime(2016, 9, 27, 23, 0, 1, 200000), 8),
    (datetime(2016, 9, 27, 23, 0, 1, 600000), 0),
    (datetime(2016, 9, 27, 23, 0, 6, 300000), 4),
])

regularized = ts.moving_average(
    start=datetime(2016, 9, 27, 23, 0, 1),
    sampling_period=1,
    placement='left',
)

这会导致:

[(datetime(2016, 9, 27, 23, 0, 1), 5.2),
 (datetime(2016, 9, 27, 23, 0, 2), 0.0),
 (datetime(2016, 9, 27, 23, 0, 3), 0.0),
 (datetime(2016, 9, 27, 23, 0, 4), 0.0),
 (datetime(2016, 9, 27, 23, 0, 5), 0.0),
 (datetime(2016, 9, 27, 23, 0, 6), 2.8)]

相关文章