您似乎在寻找.stack()。但是，您的问题似乎比 .stack() + .cumsum() 稍微复杂一些。这是因为，假设您有一个 31x12 的 DataFrame（日行、月列），您需要告诉 pandas 如何将每个月/日组合映射到一年中的某一天；否则，熊猫会假设你有 372 天年。下面的第一个函数执行此操作，下面的行首先创建一些示例数据，然后使用该函数帮助获取按年计算的累积总和。
import datetime
def stack_daily(df, year='current'):
"""Construct pd.DatetimeIndex from unstacked day/month format."""
# Confirm index is 1-indexed and ends at 31
if not np.array_equal(df.index, pd.RangeIndex(1, 32)):
raise ValueError('`df` should have `pd.RangeIndex(1, 32)`')
# Same logic for columns
if not np.array_equal(df.columns, pd.RangeIndex(1, 13)):
raise ValueError('`df` should have columns `pd.RangeIndex(1, 13)`')
if year == 'current':
year = datetime.date.today().year
stacked = df.stack() # Implicit dropna=True
day, month = zip(*stacked.index.get_values())
dates = {'year': [year] * stacked.shape[0],
'month': month,
'day': day}
return pd.to_datetime(dates)
# Create random precipitation data
import numpy as np
import pandas as pd
np.random.seed(123)
data = np.empty((31, 12))
data[:] = np.nan
mask = np.random.randint(0, 2, size=data.shape, dtype=np.bool)
vals = np.random.rand(*data.shape)
data[mask] = vals[mask]
data[29:] = np.nan
df = pd.DataFrame(data, index=pd.RangeIndex(1, 32),
columns=pd.RangeIndex(1, 13))
# Manipulate to get day-of-year index
idx = stack_daily(df)
doy = idx.dt.dayofyear
total_precip = df.stack().reset_index(drop=True)
total_precip.index = doy
total_precip.dropna(inplace=True)
total_precip = total_precip.sort_index().cumsum()
# %matplotlib inline
total_precip.plot()
更新
新功能：
import datetime
def stack_daily(df, year='current', sort=True, dropna=True):
"""Construct pd.DatetimeIndex from unstacked day/month format."""
# Confirm index is 1-indexed and ends at 31
if not np.array_equal(df.index, pd.RangeIndex(1, 32)):
raise ValueError('`df` should have `pd.RangeIndex(1, 32)`')
# Same logic for columns
if not np.array_equal(df.columns, pd.RangeIndex(1, 13)):
raise ValueError('`df` should have columns `pd.RangeIndex(1, 13)`')
if year == 'current':
year = datetime.date.today().year
stacked = df.stack(dropna=False)
year = np.repeat(year, stacked.shape[0]) # len == 372
day, month = zip(*stacked.index.get_values())
# Drop the *difference* between a valid calendar and the 372-day calendar
# created by using 12 31-day months.
# Use a pure-Python solution here because NumPy set logic doesn't generally
# support 2d arrays and we have fairly small data (1 year).
true_dates = pd.date_range(start='{}-01-01'.format(year[0]),
end='{}-12-31'.format(year[0]))
true_dates = list(zip(true_dates.day,
true_dates.month,
true_dates.year))
full_dates = list(zip(day, month, year))
# We want a boolean mask False where dates are invalid
# This should yield len(mask[mask == 1]) == 365
mask = np.array([date in true_dates for date in full_dates])
# Now filter stacked data on this mask
stacked = stacked.loc[mask]
# And finally repeat above process converting to datetime
# and then getting day of year.
day, month = zip(*stacked.index.get_values())
dates = {'year': year[:stacked.shape[0]].tolist(),
'month': month,
'day': day}
stacked.index = pd.to_datetime(dates).dt.dayofyear
if sort:
stacked.sort_index(inplace=True)
if dropna:
stacked.dropna(inplace=True)
return stacked
例子：
# Create random precipitation data
# This gets you a DataFrame with 12 months on the columns and
# 31 days on the index. Both are 1-indexed i.e. start at 1.
# There is a mix of 0.00s, NaN, and other floats, mimicking
# the data from your question.
import numpy as np
import pandas as pd
np.random.seed(123)
data = np.zeros((31, 12))
mask1 = np.random.randint(0, 2, size=data.shape, dtype=np.bool)
mask2 = np.random.randint(0, 2, size=data.shape, dtype=np.bool)
vals = np.random.rand(*data.shape)
nans = np.zeros(data.shape)
nans[:] = np.nan
data[mask1] = vals[mask1]
data[mask2] = nans[mask2]
df = pd.DataFrame(data, index=pd.RangeIndex(1, 32),
columns=pd.RangeIndex(1, 13))
# %matplotlib inline
stack_daily(df).cumsum().plot()