Numpy和Pandas插值也会更改原始数据

Question

I am trying to interpolate data for some missing days. The orginal data is;

2012-06-27 00:00:00 17
2012-06-27 01:00:00 17
2012-06-27 02:00:00 18
2012-06-27 03:00:00 18
2012-06-27 04:00:00 19
2012-06-27 05:00:00 20
2012-06-27 06:00:00 22
2012-06-27 07:00:00 23
2012-06-27 08:00:00 25
2012-06-27 09:00:00 27
2012-06-27 10:00:00 27
2012-06-27 11:00:00 29
2012-06-27 12:00:00 29
2012-06-27 13:00:00 30
2012-06-27 14:00:00 30
2012-06-27 15:00:00 29
2012-06-27 16:00:00 28
2012-06-27 17:00:00 26
2012-06-27 18:00:00 25
2012-06-27 19:00:00 24
2012-06-27 20:00:00 23
2012-06-27 21:00:00 23
2012-06-27 22:00:00 16
2012-06-27 23:00:00 15
2012-06-29 00:00:00 15
2012-06-29 01:00:00 16
2012-06-29 02:00:00 16
2012-06-29 03:00:00 16
2012-06-29 04:00:00 17
2012-06-29 05:00:00 17
2012-06-29 06:00:00 18
2012-06-29 07:00:00 19
2012-06-29 08:00:00 20
2012-06-29 09:00:00 22
2012-06-29 10:00:00 22
2012-06-29 11:00:00 22
2012-06-29 12:00:00 22
2012-06-29 13:00:00 22
2012-06-29 14:00:00 22
2012-06-29 15:00:00 22
2012-06-29 16:00:00 21
2012-06-29 17:00:00 19
2012-06-29 18:00:00 17
2012-06-29 19:00:00 16
2012-06-29 20:00:00 15
2012-06-29 21:00:00 14
2012-06-29 22:00:00 14
2012-06-29 23:00:00 13

As you can see 2014-12-28 is missing, so I tried to interpolate it using both Numpy and Pandas. For Numpy the code is;

def inter_lin_nan(ts_temp, rule):
ts_temp = ts_temp.resample(rule)
mask = np.isnan(ts_temp)
# interpolling missing values
ts_temp[mask] = np.interp(np.flatnonzero(mask), np.flatnonzero(~mask),ts_temp[~mask])
return(ts_temp)

and with Pandas I used;

df_temp=df_temp.asfreq('1h')
df_temp['Temp2'] = df_temp['temp'].interpolate(method='linear')

The problem is, both of these method does interpolate for the missing day, but they also change original data for 2014-12-29. Do you know why this is happening or am I missing something?

Answer 1

I cannot reproduce the problem, but this works for me (assuming your data frame is indexed on datetime):

df_resampled = df.resample('1H').interpolate(method='linear')

Output:

As you can see, the lines overlap perfectly for the days where there is data: no original data is 'changed'. The interpolation seems to make sense too, and in this plot the missing values in the original series were set to 0 to allow a comparison.