"""
Timeseries
----------
``TimeSeries`` is the main class in `darts`.
It represents a univariate or multivariate time series, deterministic or stochastic.
The values are stored in an array of shape `(time, dimensions, samples)`, where
`dimensions` are the dimensions (or "components", or "columns") of multivariate series,
and `samples` are samples of stochastic series.
Definitions:
- A series with `dimensions = 1` is **univariate** and a series with `dimensions > 1` is **multivariate**.
- | A series with `samples = 1` is **deterministic** and a series with `samples > 1` is
| **stochastic** (or **probabilistic**).
Each series also stores a `time_index`, which contains either datetimes (:class:`pandas.DateTimeIndex`)
or integer indices (:class:`pandas.RangeIndex`).
``TimeSeries`` are guaranteed to:
- Have a monotonically increasing time index, without holes (without missing dates)
- Contain numeric types only
- Have distinct components/columns names
- Have a well-defined frequency (`date offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_
for ``DateTimeIndex``, or step size for ``RangeIndex``)
- Have static covariates consistent with their components, or no static covariates
- Have a hierarchy consistent with their components, or no hierarchy
``TimeSeries`` can contain global or component-specific static covariate data. Static covariates in `darts` refers
to external time-invariant data that can be used by some models to help improve predictions.
Read our `user guide on covariates <https://unit8co.github.io/darts/userguide/covariates.html>`__ and the
``TimeSeries`` documentation for more information on covariates.
"""
import itertools
import pickle
import re
import sys
from collections import defaultdict
from copy import deepcopy
from inspect import signature
from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
import matplotlib.axes
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import xarray as xr
from pandas.tseries.frequencies import to_offset
from scipy.stats import kurtosis, skew
from darts.utils.utils import generate_index, n_steps_between
from .logging import get_logger, raise_if, raise_if_not, raise_log
from .utils import _build_tqdm_iterator, _parallel_apply
try:
from typing import Literal
except ImportError:
from typing_extensions import Literal
if sys.version_info >= (3, 11):
from typing import Self
else:
from typing_extensions import Self
logger = get_logger(__name__)
# dimension names in the DataArray
# the "time" one can be different, if it has a name in the underlying Series/DataFrame.
DIMS = ("time", "component", "sample")
VALID_INDEX_TYPES = (pd.DatetimeIndex, pd.RangeIndex)
STATIC_COV_TAG = "static_covariates"
DEFAULT_GLOBAL_STATIC_COV_NAME = "global_components"
HIERARCHY_TAG = "hierarchy"
[docs]class TimeSeries:
def __init__(self, xa: xr.DataArray, copy=True):
"""
Create a TimeSeries from a (well formed) DataArray.
It is recommended to use the factory methods to create TimeSeries instead.
See Also
--------
TimeSeries.from_dataframe : Create from a :class:`pandas.DataFrame`.
TimeSeries.from_group_dataframe : Create multiple TimeSeries by groups from a :class:`pandas.DataFrame`.
TimeSeries.from_series : Create from a :class:`pandas.Series`.
TimeSeries.from_values : Create from a NumPy :class:`ndarray`.
TimeSeries.from_times_and_values : Create from a time index and a Numpy :class:`ndarray`.
TimeSeries.from_csv : Create from a CSV file.
TimeSeries.from_json : Create from a JSON file.
TimeSeries.from_xarray : Create from an :class:`xarray.DataArray`.
"""
if not isinstance(xa, xr.DataArray):
raise_log(
ValueError(
"Data must be provided as an xarray DataArray instance. "
"If you need to create a TimeSeries from another type "
"(e.g. a DataFrame), look at TimeSeries factory methods "
"(e.g. TimeSeries.from_dataframe(), "
"TimeSeries.from_xarray(), TimeSeries.from_values()"
"TimeSeries.from_times_and_values(), etc...)."
),
logger,
)
if len(xa.shape) != 3:
raise_log(
ValueError(
f"TimeSeries require DataArray of dimensionality 3 ({DIMS})."
),
logger,
)
# Ideally values should be np.float, otherwise certain functionalities like diff()
# relying on np.nan (which is a float) won't work very properly.
if not np.issubdtype(xa.values.dtype, np.number):
raise_log(
ValueError("The time series must contain numeric values only."), logger
)
val_dtype = xa.values.dtype
if not (
np.issubdtype(val_dtype, np.float64) or np.issubdtype(val_dtype, np.float32)
):
logger.warning(
"TimeSeries is using a numeric type different from np.float32 or np.float64. "
"Not all functionalities may work properly. It is recommended casting your data to floating "
"point numbers before using TimeSeries."
)
if xa.dims[-2:] != DIMS[-2:]:
# The first dimension represents the time and may be named differently.
raise_log(
ValueError(
"The last two dimensions of the DataArray must be named {}".format(
DIMS[-2:]
)
),
logger,
)
# check that columns/component names are unique
components = xa.get_index(DIMS[1])
if not len(set(components)) == len(components):
raise_log(
ValueError(
f"The components (columns) names must be unique. Provided: {components}"
),
logger,
)
# how the time dimension is named; we convert hashable to string
self._time_dim = str(xa.dims[0])
# The following sorting returns a copy, which we are relying on.
# As of xarray 0.18.2, this sorting discards the freq of the index for some reason
# https://github.com/pydata/xarray/issues/5466
# We sort only if the time axis is not already sorted (monotonically increasing).
self._xa = self._sort_index(xa, copy=copy)
self._time_index = self._xa.get_index(self._time_dim)
if not isinstance(self._time_index, VALID_INDEX_TYPES):
raise_log(
ValueError(
"The time dimension of the DataArray must be indexed either with a DatetimeIndex "
"or with an RangeIndex."
),
logger,
)
self._has_datetime_index = isinstance(self._time_index, pd.DatetimeIndex)
if self._has_datetime_index:
# store original freq (see bug of sortby() above).
freq_tmp = xa.get_index(self._time_dim).freq
# if original frequency is known and positive (n > 0 -> increasing time index),
# it is guaranteed that original array was sorted and new freq must be the same.
# otherwise, infer the frequency from the sorted array
if freq_tmp is not None and freq_tmp.n > 0:
self._freq = freq_tmp
else:
self._freq = to_offset(self._xa.get_index(self._time_dim).inferred_freq)
if self._freq is None:
raise_log(
ValueError(
"The time index of the provided DataArray is missing the freq attribute, and the frequency "
"could not be directly inferred. This probably comes from inconsistent date frequencies with "
"missing dates. If you know the actual frequency, try setting `fill_missing_dates=True, "
"freq=actual_frequency`. If not, try setting `fill_missing_dates=True, freq=None` to see if a "
"frequency can be inferred."
),
logger,
)
self._freq_str: str = self._freq.freqstr
# reset freq inside the xarray index (see bug of sortby() above).
self._xa.get_index(self._time_dim).freq = self._freq
else:
self._freq: int = self._time_index.step
self._freq_str = None
# check static covariates
static_covariates = self._xa.attrs.get(STATIC_COV_TAG, None)
if not (
isinstance(static_covariates, (pd.Series, pd.DataFrame))
or static_covariates is None
):
raise_log(
ValueError(
"`static_covariates` must be either a pandas Series, DataFrame or None"
),
logger,
)
# check if valid static covariates for multivariate TimeSeries
if isinstance(static_covariates, pd.DataFrame):
n_components = len(static_covariates)
if n_components > 1 and n_components != self.n_components:
raise_log(
ValueError(
"When passing a multi-row pandas DataFrame, the number of rows must match the number of "
"components of the TimeSeries object (multi-component/multi-row static covariates "
"must map to each TimeSeries component)."
),
logger,
)
if copy:
static_covariates = static_covariates.copy()
elif isinstance(static_covariates, pd.Series):
static_covariates = static_covariates.to_frame().T
else: # None
pass
# prepare static covariates:
if static_covariates is not None:
static_covariates.index = (
self.components
if len(static_covariates) == self.n_components
else [DEFAULT_GLOBAL_STATIC_COV_NAME]
)
static_covariates.columns.name = STATIC_COV_TAG
# convert numerical columns to same dtype as series
# we get all numerical columns, except those that have right dtype already
cols_to_cast = static_covariates.select_dtypes(
include=np.number, exclude=self.dtype
).columns
# Calling astype is costly even when there's no change...
if not cols_to_cast.empty:
static_covariates = static_covariates.astype(
{col: self.dtype for col in cols_to_cast}, copy=False
)
# handle hierarchy
hierarchy = self._xa.attrs.get(HIERARCHY_TAG, None)
self._top_level_component = None
self._bottom_level_components = None
if hierarchy is not None:
if not isinstance(hierarchy, dict):
raise_log(
ValueError(
"The hierarchy must be a dict mapping (non-top) component names to their parent(s) "
"in the hierarchy."
),
logger,
)
# pre-compute grouping information
components_set = set(self.components)
children = set(hierarchy.keys())
# convert string ancestors to list of strings
hierarchy = {
k: ([v] if isinstance(v, str) else v) for k, v in hierarchy.items()
}
if not all(c in components_set for c in children):
raise_log(
ValueError(
"The keys of the hierarchy must be time series components"
),
logger,
)
ancestors = set().union(*hierarchy.values())
if not all(a in components_set for a in ancestors):
raise_log(
ValueError(
"The values of the hierarchy must only contain component names matching those "
"of the series."
),
logger,
)
hierarchy_top = components_set - children
if not len(hierarchy_top) == 1:
raise_log(
ValueError(
"The hierarchy must be such that only one component does not appear as a key "
"(the top level component)."
),
logger,
)
self._top_level_component = hierarchy_top.pop()
if self._top_level_component not in ancestors:
raise_log(
ValueError(
"Invalid hierarchy. Component {} appears as it should be top-level, but "
"does not appear as an ancestor in the hierarchy dict."
),
logger,
)
bottom_level = components_set - ancestors
# maintain the same order as the original components
self._bottom_level_components = [
c for c in self.components if c in bottom_level
]
# store static covariates and hierarchy in attributes (potentially storing None)
self._xa = _xarray_with_attrs(self._xa, static_covariates, hierarchy)
"""
Factory Methods
===============
"""
[docs] @classmethod
def from_xarray(
cls,
xa: xr.DataArray,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
fillna_value: Optional[float] = None,
) -> Self:
"""
Return a TimeSeries instance built from an xarray DataArray.
The dimensions of the DataArray have to be (time, component, sample), in this order. The time
dimension can have an arbitrary name, but component and sample must be named "component" and "sample",
respectively.
The first dimension (time), and second dimension (component) must be indexed (i.e., have coordinates).
The time must be indexed either with a pandas DatetimeIndex, a pandas RangeIndex, or a pandas Index that can
be converted to a RangeIndex. It is better if the index has no holes; alternatively setting
`fill_missing_dates` can in some cases solve these issues (filling holes with NaN, or with the provided
`fillna_value` numeric value, if any).
If two components have the same name or are not strings, this method will disambiguate the components
names by appending a suffix of the form "<name>_N" to the N-th column with name "name".
The component names in the static covariates and hierarchy (if any) are *not* disambiguated.
Parameters
----------
xa
The xarray DataArray
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
Returns
-------
TimeSeries
A univariate or multivariate deterministic TimeSeries constructed from the inputs.
"""
xa_index = xa.get_index(xa.dims[0])
has_datetime_index = isinstance(xa_index, pd.DatetimeIndex)
has_range_index = isinstance(xa_index, pd.RangeIndex)
has_integer_index = not (has_datetime_index or has_range_index)
has_frequency = (
has_datetime_index and xa_index.freq is not None
) or has_range_index
# optionally fill missing dates; do it only when there is a DatetimeIndex (and not a RangeIndex)
if fill_missing_dates:
xa_ = cls._fill_missing_dates(xa, freq=freq)
# provided index does not have a freq; using the provided freq
elif (
(has_datetime_index or has_integer_index)
and freq is not None
and not has_frequency
):
xa_ = cls._restore_xarray_from_frequency(xa, freq=freq)
# index is an integer index and no freq is provided; try convert it to pd.RangeIndex
elif has_integer_index and freq is None:
xa_ = cls._integer_to_range_indexed_xarray(xa)
else:
xa_ = xa
if fillna_value is not None:
xa_ = xa_.fillna(fillna_value)
# clean components (columns) names if needed (if names are not unique, or not strings)
components = xa_.get_index(DIMS[1])
if len(set(components)) != len(components) or any(
[not isinstance(s, str) for s in components]
):
def _clean_component_list(columns) -> List[str]:
# return a list of string containing column names
# make each column name unique in case some columns have the same names
clist = columns.to_list()
# convert everything to string if needed
for i, column in enumerate(clist):
if not isinstance(column, str):
clist[i] = str(column)
has_duplicate = len(set(clist)) != len(clist)
while has_duplicate:
# we may have to loop several times (e.g. we could have columns ["0", "0_1", "0"] and not
# noticing when renaming the last "0" into "0_1" that "0_1" already exists...)
name_to_occurence = defaultdict(int)
for i, column in enumerate(clist):
name_to_occurence[clist[i]] += 1
if name_to_occurence[clist[i]] > 1:
clist[i] = clist[i] + "_{}".format(
name_to_occurence[clist[i]] - 1
)
has_duplicate = len(set(clist)) != len(clist)
return clist
time_index_name = xa_.dims[0]
columns_list = _clean_component_list(components)
# Note: an option here could be to also rename the component names in the static covariates
# and/or hierarchy, if any. However, we decide not to do so as those are directly dependent on the
# component names to work properly, so in case there's any name conflict it's better solved
# by the user than handled by silent renaming, which can change the way things work.
# TODO: is there a way to just update the component index without re-creating a new DataArray?
# -> Answer: Yes, but it's slower: e.g.:
# ```
# xa_ = xa_.assign_coords(
# {
# time_index_name: xa_.get_index(time_index_name),
# DIMS[1]: columns_list
# }
# )
# ```
xa_ = xr.DataArray(
xa_.values,
dims=xa_.dims,
coords={
time_index_name: xa_.get_index(time_index_name),
DIMS[1]: columns_list,
},
attrs=xa_.attrs,
)
# We cast the array to float
if np.issubdtype(xa_.values.dtype, np.float32) or np.issubdtype(
xa_.values.dtype, np.float64
):
return cls(xa_)
else:
return cls(xa_.astype(np.float64))
[docs] @classmethod
def from_csv(
cls,
filepath_or_buffer,
time_col: Optional[str] = None,
value_cols: Optional[Union[List[str], str]] = None,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
fillna_value: Optional[float] = None,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
hierarchy: Optional[Dict] = None,
**kwargs,
) -> Self:
"""
Build a deterministic TimeSeries instance built from a single CSV file.
One column can be used to represent the time (if not present, the time index will be a RangeIndex)
and a list of columns `value_cols` can be used to indicate the values for this time series.
Parameters
----------
filepath_or_buffer
The path to the CSV file, or the file object; consistent with the argument of `pandas.read_csv` function
time_col
The time column name. If set, the column will be cast to a pandas DatetimeIndex (if it contains
timestamps) or a RangeIndex (if it contains integers).
If not set, the pandas RangeIndex will be used.
value_cols
A string or list of strings representing the value column(s) to be extracted from the CSV file. If set to
`None`, all columns from the CSV file will be used (except for the time_col, if specified)
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a pandas
DataFrame. If a Series, the index represents the static variables. The covariates are globally 'applied'
to all components of the TimeSeries. If a DataFrame, the columns represent the static variables and the
rows represent the components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates
are globally 'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of
rows must match the number of components of the TimeSeries (in this case, the number of columns in the CSV
file). This adds control for component-specific static covariates.
hierarchy
Optionally, a dictionary describing the grouping(s) of the time series. The keys are component names, and
for a given component name `c`, the value is a list of component names that `c` "belongs" to. For instance,
if there is a `total` component, split both in two divisions `d1` and `d2` and in two regions `r1` and `r2`,
and four products `d1r1` (in division `d1` and region `r1`), `d2r1`, `d1r2` and `d2r2`, the hierarchy would
be encoded as follows.
.. highlight:: python
.. code-block:: python
hierarchy={
"d1r1": ["d1", "r1"],
"d1r2": ["d1", "r2"],
"d2r1": ["d2", "r1"],
"d2r2": ["d2", "r2"],
"d1": ["total"],
"d2": ["total"],
"r1": ["total"],
"r2": ["total"]
}
..
The hierarchy can be used to reconcile forecasts (so that the sums of the forecasts at
different levels are consistent), see `hierarchical reconciliation
<https://unit8co.github.io/darts/generated_api/darts.dataprocessing.transformers.reconciliation.html>`_.
**kwargs
Optional arguments to be passed to `pandas.read_csv` function
Returns
-------
TimeSeries
A univariate or multivariate deterministic TimeSeries constructed from the inputs.
"""
df = pd.read_csv(filepath_or_buffer=filepath_or_buffer, **kwargs)
return cls.from_dataframe(
df=df,
time_col=time_col,
value_cols=value_cols,
fill_missing_dates=fill_missing_dates,
freq=freq,
fillna_value=fillna_value,
static_covariates=static_covariates,
hierarchy=hierarchy,
)
[docs] @classmethod
def from_dataframe(
cls,
df: pd.DataFrame,
time_col: Optional[str] = None,
value_cols: Optional[Union[List[str], str]] = None,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
fillna_value: Optional[float] = None,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
hierarchy: Optional[Dict] = None,
) -> Self:
"""
Build a deterministic TimeSeries instance built from a selection of columns of a DataFrame.
One column (or the DataFrame index) has to represent the time,
and a list of columns `value_cols` has to represent the values for this time series.
Parameters
----------
df
The DataFrame
time_col
The time column name. If set, the column will be cast to a pandas DatetimeIndex (if it contains
timestamps) or a RangeIndex (if it contains integers).
If not set, the DataFrame index will be used. In this case the DataFrame must contain an index that is
either a pandas DatetimeIndex, a pandas RangeIndex, or a pandas Index that can be converted to a
RangeIndex. It is better if the index has no holes; alternatively setting `fill_missing_dates` can in some
cases solve these issues (filling holes with NaN, or with the provided `fillna_value` numeric value, if
any).
value_cols
A string or list of strings representing the value column(s) to be extracted from the DataFrame. If set to
`None`, the whole DataFrame will be used.
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a pandas
DataFrame. If a Series, the index represents the static variables. The covariates are globally 'applied'
to all components of the TimeSeries. If a DataFrame, the columns represent the static variables and the
rows represent the components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates
are globally 'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of
rows must match the number of components of the TimeSeries (in this case, the number of columns in
``value_cols``). This adds control for component-specific static covariates.
hierarchy
Optionally, a dictionary describing the grouping(s) of the time series. The keys are component names, and
for a given component name `c`, the value is a list of component names that `c` "belongs" to. For instance,
if there is a `total` component, split both in two divisions `d1` and `d2` and in two regions `r1` and `r2`,
and four products `d1r1` (in division `d1` and region `r1`), `d2r1`, `d1r2` and `d2r2`, the hierarchy would
be encoded as follows.
.. highlight:: python
.. code-block:: python
hierarchy={
"d1r1": ["d1", "r1"],
"d1r2": ["d1", "r2"],
"d2r1": ["d2", "r1"],
"d2r2": ["d2", "r2"],
"d1": ["total"],
"d2": ["total"],
"r1": ["total"],
"r2": ["total"]
}
..
The hierarchy can be used to reconcile forecasts (so that the sums of the forecasts at
different levels are consistent), see `hierarchical reconciliation
<https://unit8co.github.io/darts/generated_api/darts.dataprocessing.transformers.reconciliation.html>`_.
Returns
-------
TimeSeries
A univariate or multivariate deterministic TimeSeries constructed from the inputs.
"""
# get values
if value_cols is None:
series_df = df.loc[:, df.columns != time_col]
else:
if isinstance(value_cols, str):
value_cols = [value_cols]
series_df = df[value_cols]
# get time index
if time_col:
if time_col not in df.columns:
raise_log(AttributeError(f"time_col='{time_col}' is not present."))
time_index = pd.Index([])
time_col_vals = df[time_col]
if np.issubdtype(time_col_vals.dtype, object):
# Try to convert to integers if needed
try:
time_col_vals = time_col_vals.astype(int)
except ValueError:
pass
if np.issubdtype(time_col_vals.dtype, np.integer):
# We have to check all integers appear only once to have a valid index
raise_if(
time_col_vals.duplicated().any(),
"The provided integer time index column contains duplicate values.",
)
# Temporarily use an integer Index to sort the values, and replace by a
# RangeIndex in `TimeSeries.from_xarray()`
time_index = pd.Index(time_col_vals)
elif np.issubdtype(time_col_vals.dtype, object):
# The integer conversion failed; try datetimes
try:
time_index = pd.DatetimeIndex(time_col_vals)
except ValueError:
raise_log(
AttributeError(
"'time_col' is of 'object' dtype but doesn't contain valid timestamps"
)
)
elif np.issubdtype(time_col_vals.dtype, np.datetime64):
time_index = pd.DatetimeIndex(time_col_vals)
else:
raise_log(
AttributeError(
"Invalid type of `time_col`: it needs to be of either 'str', 'datetime' or 'int' dtype."
)
)
time_index.name = time_col
else:
raise_if_not(
isinstance(df.index, VALID_INDEX_TYPES)
or np.issubdtype(df.index.dtype, np.integer),
"If time_col is not specified, the DataFrame must be indexed either with "
"a DatetimeIndex, a RangeIndex, or an integer Index that can be converted into a RangeIndex",
logger,
)
# BUGFIX : force time-index to be timezone naive as xarray doesn't support it
# pandas.DataFrame loses the tz information if it's not its index
if isinstance(df.index, pd.DatetimeIndex) and df.index.tz is not None:
logger.warning(
"The provided DatetimeIndex was associated with a timezone, which is currently not supported "
"by xarray. To avoid unexpected behaviour, the tz information was removed. Consider calling "
f"`ts.time_index.tz_localize({df.index.tz})` when exporting the results."
"To plot the series with the right time steps, consider setting the matplotlib.pyplot "
"`rcParams['timezone']` parameter to automatically convert the time axis back to the "
"original timezone."
)
time_index = df.index.tz_localize(None)
else:
time_index = df.index
if not time_index.name:
time_index.name = time_col if time_col else DIMS[0]
if series_df.columns.name:
series_df.columns.name = None
xa = xr.DataArray(
series_df.values[:, :, np.newaxis],
dims=(time_index.name,) + DIMS[-2:],
coords={time_index.name: time_index, DIMS[1]: series_df.columns},
attrs={STATIC_COV_TAG: static_covariates, HIERARCHY_TAG: hierarchy},
)
return cls.from_xarray(
xa=xa,
fill_missing_dates=fill_missing_dates,
freq=freq,
fillna_value=fillna_value,
)
[docs] @classmethod
def from_group_dataframe(
cls,
df: pd.DataFrame,
group_cols: Union[List[str], str],
time_col: Optional[str] = None,
value_cols: Optional[Union[List[str], str]] = None,
static_cols: Optional[Union[List[str], str]] = None,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
fillna_value: Optional[float] = None,
drop_group_cols: Optional[Union[List[str], str]] = None,
n_jobs: Optional[int] = 1,
verbose: Optional[bool] = False,
) -> List[Self]:
"""
Build a list of TimeSeries instances grouped by a selection of columns from a DataFrame.
One column (or the DataFrame index) has to represent the time,
a list of columns `group_cols` must be used for extracting the individual TimeSeries by groups,
and a list of columns `value_cols` has to represent the values for the individual time series.
Values from columns ``group_cols`` and ``static_cols`` are added as static covariates to the resulting
TimeSeries objects. These can be viewed with `my_series.static_covariates`. Different to `group_cols`,
`static_cols` only adds the static values but are not used to extract the TimeSeries groups.
Parameters
----------
df
The DataFrame
group_cols
A string or list of strings representing the columns from the DataFrame by which to extract the
individual TimeSeries groups.
time_col
The time column name. If set, the column will be cast to a pandas DatetimeIndex (if it contains
timestamps) or a RangeIndex (if it contains integers).
If not set, the DataFrame index will be used. In this case the DataFrame must contain an index that is
either a pandas DatetimeIndex, a pandas RangeIndex, or a pandas Index that can be converted to a
RangeIndex. Be aware that the index must represents the actual index of each individual time series group
(can contain non-unique values). It is better if the index has no holes; alternatively setting
`fill_missing_dates` can in some cases solve these issues (filling holes with NaN, or with the provided
`fillna_value` numeric value, if any).
value_cols
A string or list of strings representing the value column(s) to be extracted from the DataFrame. If set to
`None`, the whole DataFrame will be used.
static_cols
A string or list of strings representing static variable columns from the DataFrame that should be
appended as static covariates to the resulting TimeSeries groups. Different to `group_cols`, the
DataFrame is not grouped by these columns. Note that for every group, there must be exactly one
unique value.
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
drop_group_cols
Optionally, a string or list of strings with `group_cols` column(s) to exclude from the static covariates.
n_jobs
Optionally, an integer representing the number of parallel jobs to run. Behavior is the same as in the
`joblib.Parallel` class.
verbose
Optionally, a boolean value indicating whether to display a progress bar.
Returns
-------
List[TimeSeries]
A list containing a univariate or multivariate deterministic TimeSeries per group in the DataFrame.
"""
if time_col is None and df.index.is_monotonic_increasing:
logger.warning(
"UserWarning: `time_col` was not set and `df` has a monotonically increasing (time) index. This "
"results in time series groups with non-overlapping (time) index. You can ignore this warning if the "
"index represents the actual index of each individual time series group."
)
group_cols = [group_cols] if not isinstance(group_cols, list) else group_cols
if drop_group_cols:
drop_group_cols = (
[drop_group_cols]
if not isinstance(drop_group_cols, list)
else drop_group_cols
)
invalid_cols = set(drop_group_cols) - set(group_cols)
if invalid_cols:
raise_log(
ValueError(
f"Found invalid `drop_group_cols` columns. All columns must be in the passed `group_cols`. "
f"Expected any of: {group_cols}, received: {invalid_cols}."
),
logger=logger,
)
drop_group_col_idx = [
idx for idx, col in enumerate(group_cols) if col in drop_group_cols
]
else:
drop_group_cols = []
drop_group_col_idx = []
if static_cols is not None:
static_cols = (
[static_cols] if not isinstance(static_cols, list) else static_cols
)
else:
static_cols = []
static_cov_cols = group_cols + static_cols
extract_static_cov_cols = [
col for col in static_cov_cols if col not in drop_group_cols
]
extract_time_col = [] if time_col is None else [time_col]
if value_cols is None:
value_cols = df.columns.drop(static_cov_cols + extract_time_col).tolist()
extract_value_cols = [value_cols] if isinstance(value_cols, str) else value_cols
df = df[static_cov_cols + extract_value_cols + extract_time_col]
# sort on entire `df` to avoid having to sort individually later on
if time_col:
df.index = pd.DatetimeIndex(df[time_col])
df = df.drop(columns=time_col)
df = df.sort_index()
groups = df.groupby(group_cols[0] if len(group_cols) == 1 else group_cols)
iterator = _build_tqdm_iterator(
groups,
verbose=verbose,
total=len(groups),
desc="Creating TimeSeries",
)
def from_group(static_cov_vals, group):
split = group[extract_value_cols]
static_cov_vals = (
(static_cov_vals,)
if not isinstance(static_cov_vals, tuple)
else static_cov_vals
)
# optionally, exclude group columns from static covariates
if drop_group_col_idx:
if len(drop_group_col_idx) == len(group_cols):
static_cov_vals = tuple()
else:
static_cov_vals = tuple(
val
for idx, val in enumerate(static_cov_vals)
if idx not in drop_group_col_idx
)
# check that for each group there is only one unique value per column in `static_cols`
if static_cols:
static_cols_valid = [
len(group[col].unique()) == 1 for col in static_cols
]
if not all(static_cols_valid):
# encountered performance issues when evaluating the error message from below in every
# iteration with `raise_if_not(all(static_cols_valid), message, logger)`
invalid_cols = [
static_col
for static_col, is_valid in zip(static_cols, static_cols_valid)
if not is_valid
]
raise_if(
True,
f"Encountered more than one unique value in group {group} for given static columns: "
f"{invalid_cols}.",
logger,
)
# add the static covariates to the group values
static_cov_vals += tuple(group[static_cols].values[0])
return cls.from_dataframe(
df=split,
fill_missing_dates=fill_missing_dates,
freq=freq,
fillna_value=fillna_value,
static_covariates=(
pd.DataFrame([static_cov_vals], columns=extract_static_cov_cols)
if extract_static_cov_cols
else None
),
)
return _parallel_apply(
iterator,
from_group,
n_jobs,
fn_args=dict(),
fn_kwargs=dict(),
)
[docs] @classmethod
def from_series(
cls,
pd_series: pd.Series,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
fillna_value: Optional[float] = None,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
) -> Self:
"""
Build a univariate deterministic series from a pandas Series.
The series must contain an index that is either a pandas DatetimeIndex, a pandas RangeIndex, or a pandas Index
that can be converted into a RangeIndex. It is better if the index has no holes; alternatively setting
`fill_missing_dates` can in some cases solve these issues (filling holes with NaN, or with the provided
`fillna_value` numeric value, if any).
Parameters
----------
pd_series
The pandas Series instance.
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a
single-row pandas DataFrame. If a Series, the index represents the static variables. If a DataFrame, the
columns represent the static variables and the single row represents the univariate TimeSeries component.
Returns
-------
TimeSeries
A univariate and deterministic TimeSeries constructed from the inputs.
"""
df = pd.DataFrame(pd_series)
return cls.from_dataframe(
df,
time_col=None,
value_cols=None,
fill_missing_dates=fill_missing_dates,
freq=freq,
fillna_value=fillna_value,
static_covariates=static_covariates,
)
[docs] @classmethod
def from_times_and_values(
cls,
times: Union[pd.DatetimeIndex, pd.RangeIndex, pd.Index],
values: np.ndarray,
fill_missing_dates: Optional[bool] = False,
freq: Optional[Union[str, int]] = None,
columns: Optional[pd._typing.Axes] = None,
fillna_value: Optional[float] = None,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
hierarchy: Optional[Dict] = None,
) -> Self:
"""
Build a series from a time index and value array.
Parameters
----------
times
A pandas DateTimeIndex, RangeIndex, or Index that can be converted to a RangeIndex representing the time
axis for the time series. It is better if the index has no holes; alternatively setting
`fill_missing_dates` can in some cases solve these issues (filling holes with NaN, or with the provided
`fillna_value` numeric value, if any).
values
A Numpy array of values for the TimeSeries. Both 2-dimensional arrays, for deterministic series,
and 3-dimensional arrays, for probabilistic series, are accepted. In the former case the dimensions
should be (time, component), and in the latter case (time, component, sample).
fill_missing_dates
Optionally, a boolean value indicating whether to fill missing dates (or indices in case of integer index)
with NaN values. This requires either a provided `freq` or the possibility to infer the frequency from the
provided timestamps. See :meth:`_fill_missing_dates() <TimeSeries._fill_missing_dates>` for more info.
freq
Optionally, a string or integer representing the frequency of the underlying index. This is useful in order
to fill in missing values if some dates are missing and `fill_missing_dates` is set to `True`.
If a string, represents the frequency of the pandas DatetimeIndex (see `offset aliases
<https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_ for more info on
supported frequencies).
If an integer, represents the step size of the pandas Index or pandas RangeIndex.
columns
Columns to be used by the underlying pandas DataFrame.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a pandas
DataFrame. If a Series, the index represents the static variables. The covariates are globally 'applied'
to all components of the TimeSeries. If a DataFrame, the columns represent the static variables and the
rows represent the components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates
are globally 'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of
rows must match the number of components of the TimeSeries (in this case, the number of columns in
``values``). This adds control for component-specific static covariates.
hierarchy
Optionally, a dictionary describing the grouping(s) of the time series. The keys are component names, and
for a given component name `c`, the value is a list of component names that `c` "belongs" to. For instance,
if there is a `total` component, split both in two divisions `d1` and `d2` and in two regions `r1` and `r2`,
and four products `d1r1` (in division `d1` and region `r1`), `d2r1`, `d1r2` and `d2r2`, the hierarchy would
be encoded as follows.
.. highlight:: python
.. code-block:: python
hierarchy={
"d1r1": ["d1", "r1"],
"d1r2": ["d1", "r2"],
"d2r1": ["d2", "r1"],
"d2r2": ["d2", "r2"],
"d1": ["total"],
"d2": ["total"],
"r1": ["total"],
"r2": ["total"]
}
..
The hierarchy can be used to reconcile forecasts (so that the sums of the forecasts at
different levels are consistent), see `hierarchical reconciliation
<https://unit8co.github.io/darts/generated_api/darts.dataprocessing.transformers.reconciliation.html>`_.
Returns
-------
TimeSeries
A TimeSeries constructed from the inputs.
"""
raise_if_not(
isinstance(times, VALID_INDEX_TYPES)
or np.issubdtype(times.dtype, np.integer),
"the `times` argument must be a RangeIndex, or a DateTimeIndex. Use "
"TimeSeries.from_values() if you want to use an automatic RangeIndex.",
)
# BUGFIX : force time-index to be timezone naive as xarray doesn't support it
if isinstance(times, pd.DatetimeIndex) and times.tz is not None:
logger.warning(
"The `times` argument was associated with a timezone, which is currently not supported "
"by xarray. To avoid unexpected behaviour, the tz information was removed. Consider calling "
f"`ts.time_index.tz_localize({times.tz})` when exporting the results."
"To plot the series with the right time steps, consider setting the matplotlib.pyplot "
"`rcParams['timezone']` parameter to automatically convert the time axis back to the "
"original timezone."
)
times = times.tz_localize(None)
times_name = DIMS[0] if not times.name else times.name
# avoid copying if data is already np.ndarray:
values = np.array(values) if not isinstance(values, np.ndarray) else values
if len(values.shape) == 1:
values = np.expand_dims(values, 1)
if len(values.shape) == 2:
values = np.expand_dims(values, 2)
coords = {times_name: times}
if columns is not None:
coords[DIMS[1]] = columns
xa = xr.DataArray(
values,
dims=(times_name,) + DIMS[-2:],
coords=coords,
attrs={STATIC_COV_TAG: static_covariates, HIERARCHY_TAG: hierarchy},
)
return cls.from_xarray(
xa=xa,
fill_missing_dates=fill_missing_dates,
freq=freq,
fillna_value=fillna_value,
)
[docs] @classmethod
def from_values(
cls,
values: np.ndarray,
columns: Optional[pd._typing.Axes] = None,
fillna_value: Optional[float] = None,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
hierarchy: Optional[Dict] = None,
) -> Self:
"""
Build an integer-indexed series from an array of values.
The series will have an integer index (RangeIndex).
Parameters
----------
values
A Numpy array of values for the TimeSeries. Both 2-dimensional arrays, for deterministic series,
and 3-dimensional arrays, for probabilistic series, are accepted. In the former case the dimensions
should be (time, component), and in the latter case (time, component, sample).
columns
Columns to be used by the underlying pandas DataFrame.
fillna_value
Optionally, a numeric value to fill missing values (NaNs) with.
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a pandas
DataFrame. If a Series, the index represents the static variables. The covariates are globally 'applied'
to all components of the TimeSeries. If a DataFrame, the columns represent the static variables and the
rows represent the components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates
are globally 'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of
rows must match the number of components of the TimeSeries (in this case, the number of columns in
``values``). This adds control for component-specific static covariates.
hierarchy
Optionally, a dictionary describing the grouping(s) of the time series. The keys are component names, and
for a given component name `c`, the value is a list of component names that `c` "belongs" to. For instance,
if there is a `total` component, split both in two divisions `d1` and `d2` and in two regions `r1` and `r2`,
and four products `d1r1` (in division `d1` and region `r1`), `d2r1`, `d1r2` and `d2r2`, the hierarchy would
be encoded as follows.
.. highlight:: python
.. code-block:: python
hierarchy={
"d1r1": ["d1", "r1"],
"d1r2": ["d1", "r2"],
"d2r1": ["d2", "r1"],
"d2r2": ["d2", "r2"],
"d1": ["total"],
"d2": ["total"],
"r1": ["total"],
"r2": ["total"]
}
..
The hierarchy can be used to reconcile forecasts (so that the sums of the forecasts at
different levels are consistent), see `hierarchical reconciliation
<https://unit8co.github.io/darts/generated_api/darts.dataprocessing.transformers.reconciliation.html>`_.
Returns
-------
TimeSeries
A TimeSeries constructed from the inputs.
"""
time_index = pd.RangeIndex(0, len(values), 1)
values_ = (
np.reshape(values, (len(values), 1)) if len(values.shape) == 1 else values
)
return cls.from_times_and_values(
times=time_index,
values=values_,
fill_missing_dates=False,
freq=None,
columns=columns,
fillna_value=fillna_value,
static_covariates=static_covariates,
hierarchy=hierarchy,
)
[docs] @classmethod
def from_json(
cls,
json_str: str,
static_covariates: Optional[Union[pd.Series, pd.DataFrame]] = None,
hierarchy: Optional[Dict] = None,
) -> Self:
"""
Build a series from the JSON String representation of a ``TimeSeries``
(produced using :func:`TimeSeries.to_json()`).
At the moment this only supports deterministic time series (i.e., made of 1 sample).
Parameters
----------
json_str
The JSON String to convert
static_covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series or a pandas
DataFrame. If a Series, the index represents the static variables. The covariates are globally 'applied'
to all components of the TimeSeries. If a DataFrame, the columns represent the static variables and the
rows represent the components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates
are globally 'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of
rows must match the number of components of the TimeSeries (in this case, the number of columns in
``value_cols``). This adds control for component-specific static covariates.
hierarchy
Optionally, a dictionary describing the grouping(s) of the time series. The keys are component names, and
for a given component name `c`, the value is a list of component names that `c` "belongs" to. For instance,
if there is a `total` component, split both in two divisions `d1` and `d2` and in two regions `r1` and `r2`,
and four products `d1r1` (in division `d1` and region `r1`), `d2r1`, `d1r2` and `d2r2`, the hierarchy would
be encoded as follows.
.. highlight:: python
.. code-block:: python
hierarchy={
"d1r1": ["d1", "r1"],
"d1r2": ["d1", "r2"],
"d2r1": ["d2", "r1"],
"d2r2": ["d2", "r2"],
"d1": ["total"],
"d2": ["total"],
"r1": ["total"],
"r2": ["total"]
}
..
The hierarchy can be used to reconcile forecasts (so that the sums of the forecasts at
different levels are consistent), see `hierarchical reconciliation
<https://unit8co.github.io/darts/generated_api/darts.dataprocessing.transformers.reconciliation.html>`_.
Returns
-------
TimeSeries
The time series object converted from the JSON String
"""
df = pd.read_json(json_str, orient="split")
return cls.from_dataframe(
df, static_covariates=static_covariates, hierarchy=hierarchy
)
[docs] @classmethod
def from_pickle(cls, path: str) -> Self:
"""
Read a pickled ``TimeSeries``.
Parameters
----------
path : string
path pointing to a pickle file that will be loaded
Returns
-------
TimeSeries
timeseries object loaded from file
Notes
-----
Xarray docs [1]_ suggest not using pickle as a long-term data storage.
References
----------
.. [1] http://xarray.pydata.org/en/stable/user-guide/io.html#pickle
"""
with open(path, "rb") as fh:
return pickle.load(fh)
"""
Properties
==========
"""
@property
def static_covariates(self) -> Optional[pd.DataFrame]:
"""
Returns the static covariates contained in the series as a pandas DataFrame.
The columns represent the static variables and the rows represent the components of the uni/multivariate
series.
"""
return self._xa.attrs.get(STATIC_COV_TAG, None)
@property
def hierarchy(self) -> Optional[Dict]:
"""
The hierarchy of this TimeSeries, if any.
If set, the hierarchy is encoded as a dictionary, whose keys are individual components
and values are the set of parent(s) of these components in the hierarchy.
"""
return self._xa.attrs.get(HIERARCHY_TAG, None)
@property
def has_hierarchy(self) -> bool:
"""Whether this series is hierarchical or not."""
return self.hierarchy is not None
@property
def top_level_component(self) -> Optional[str]:
"""
The top level component name of this series, or None if the series has no hierarchy.
"""
return self._top_level_component
@property
def bottom_level_components(self) -> Optional[List[str]]:
"""
The bottom level component names of this series, or None if the series has no hierarchy.
"""
return self._bottom_level_components
@property
def top_level_series(self) -> Optional[Self]:
"""
The univariate series containing the single top-level component of this series,
or None if the series has no hierarchy.
"""
return self[self.top_level_component] if self.has_hierarchy else None
@property
def bottom_level_series(self) -> Optional[List[Self]]:
"""
The series containing the bottom-level components of this series in the same
order as they appear in the series, or None if the series has no hierarchy.
The returned series is multivariate if there are multiple bottom components.
"""
return (
self[[c for c in self.components if c in self.bottom_level_components]]
if self.has_hierarchy
else None
)
@property
def n_samples(self) -> int:
"""Number of samples contained in the series."""
return len(self._xa.sample)
@property
def n_components(self) -> int:
"""Number of components (dimensions) contained in the series."""
return len(self._xa.component)
@property
def width(self) -> int:
""" "Width" (= number of components) of the series."""
return self.n_components
@property
def n_timesteps(self) -> int:
"""Number of time steps in the series."""
return len(self._time_index)
@property
def is_deterministic(self) -> bool:
"""Whether this series is deterministic."""
return self.n_samples == 1
@property
def is_stochastic(self) -> bool:
"""Whether this series is stochastic."""
return not self.is_deterministic
@property
def is_probabilistic(self) -> bool:
"""Whether this series is stochastic (= probabilistic)."""
return self.is_stochastic
@property
def is_univariate(self) -> bool:
"""Whether this series is univariate."""
return self.n_components == 1
@property
def freq(self) -> Union[pd.DateOffset, int]:
"""The frequency of the series.
A `pd.DateOffset` if series is indexed with a `pd.DatetimeIndex`.
An integer (step size) if series is indexed with a `pd.RangeIndex`.
"""
return self._freq
@property
def freq_str(self) -> str:
"""The frequency string representation of the series."""
return self._freq_str
@property
def dtype(self):
"""The dtype of the series' values."""
return self._xa.values.dtype
@property
def components(self) -> pd.Index:
"""The names of the components, as a Pandas Index."""
return self._xa.get_index(DIMS[1]).copy()
@property
def columns(self) -> pd.Index:
"""The names of the components, as a Pandas Index."""
return self.components
@property
def time_index(self) -> Union[pd.DatetimeIndex, pd.RangeIndex]:
"""The time index of this time series."""
return self._time_index.copy()
@property
def time_dim(self) -> str:
"""The name of the time dimension for this time series."""
return self._time_dim
@property
def has_datetime_index(self) -> bool:
"""Whether this series is indexed with a DatetimeIndex (otherwise it is indexed with an RangeIndex)."""
return self._has_datetime_index
@property
def has_range_index(self) -> bool:
"""Whether this series is indexed with an RangeIndex (otherwise it is indexed with a DatetimeIndex)."""
return not self._has_datetime_index
@property
def has_static_covariates(self) -> bool:
"""Whether this series contains static covariates."""
return self.static_covariates is not None
@property
def duration(self) -> Union[pd.Timedelta, int]:
"""The duration of this time series (as a time delta or int)."""
return self._time_index[-1] - self._time_index[0]
"""
Some asserts
=============
"""
# TODO: put at the bottom
def _assert_univariate(self):
if not self.is_univariate:
raise_log(
AssertionError(
"Only univariate TimeSeries instances support this method"
),
logger,
)
def _assert_deterministic(self):
if not self.is_deterministic:
raise_log(
AssertionError(
"Only deterministic TimeSeries (with 1 sample) instances support this method"
),
logger,
)
def _assert_stochastic(self):
if not self.is_stochastic:
raise_log(
AssertionError(
"Only non-deterministic TimeSeries (with more than 1 samples) "
"instances support this method"
),
logger,
)
def _raise_if_not_within(self, ts: Union[pd.Timestamp, int]):
if isinstance(ts, pd.Timestamp):
# Not that the converse doesn't apply (a time-indexed series can be called with an integer)
raise_if_not(
self._has_datetime_index,
"Function called with a timestamp, but series not time-indexed.",
logger,
)
is_inside = self.start_time() <= ts <= self.end_time()
else:
if self._has_datetime_index:
is_inside = 0 <= ts <= len(self)
else:
is_inside = self.start_time() <= ts <= self.end_time()
raise_if_not(
is_inside,
"Timestamp must be between {} and {}".format(
self.start_time(), self.end_time()
),
logger,
)
def _get_first_timestamp_after(self, ts: pd.Timestamp) -> Union[pd.Timestamp, int]:
return next(filter(lambda t: t >= ts, self._time_index))
def _get_last_timestamp_before(self, ts: pd.Timestamp) -> Union[pd.Timestamp, int]:
return next(filter(lambda t: t <= ts, self._time_index[::-1]))
"""
Export functions
================
"""
[docs] def data_array(self, copy=True) -> xr.DataArray:
"""
Return the ``xarray.DataArray`` representation underlying this series.
Parameters
----------
copy
Whether to return a copy of the series. Leave it to True unless you know what you are doing.
Returns
-------
xarray.DataArray
The xarray DataArray underlying this time series.
"""
return self._xa.copy() if copy else self._xa
[docs] def pd_series(self, copy=True) -> pd.Series:
"""
Return a Pandas Series representation of this univariate deterministic time series.
Works only for univariate series that are deterministic (i.e., made of 1 sample).
Parameters
----------
copy
Whether to return a copy of the series. Leave it to True unless you know what you are doing.
Returns
-------
pandas.Series
A Pandas Series representation of this univariate time series.
"""
self._assert_univariate()
self._assert_deterministic()
if copy:
return pd.Series(
self._xa[:, 0, 0].values.copy(),
index=self._time_index.copy(),
name=self.components[0],
)
else:
return pd.Series(
self._xa[:, 0, 0].values,
index=self._time_index,
name=self.components[0],
)
[docs] def pd_dataframe(self, copy=True, suppress_warnings=False) -> pd.DataFrame:
"""
Return a Pandas DataFrame representation of this time series.
Each of the series components will appear as a column in the DataFrame.
If the series is stochastic, the samples are returned as columns of the dataframe with column names
as 'component_s#' (e.g. with two components and two samples:
'comp0_s0', 'comp0_s1' 'comp1_s0' 'comp1_s1').
Parameters
----------
copy
Whether to return a copy of the dataframe. Leave it to True unless you know what you are doing.
Returns
-------
pandas.DataFrame
The Pandas DataFrame representation of this time series
"""
if not self.is_deterministic:
if not suppress_warnings:
logger.warning(
"You are transforming a stochastic TimeSeries (i.e., contains several samples). "
"The resulting DataFrame is a 2D object with all samples on the columns. "
"If this is not the expected behavior consider calling a function "
"adapted to stochastic TimeSeries like quantile_df()."
)
comp_name = list(self.components)
samples = range(self.n_samples)
df_col_names = [
"_s".join((comp_name, str(sample_id)))
for comp_name, sample_id in itertools.product(comp_name, samples)
]
if copy:
return pd.DataFrame(
self._xa.stack(data=(DIMS[1], DIMS[2])).values.copy(),
index=self._time_index.copy(),
columns=df_col_names.copy(),
)
else:
return pd.DataFrame(
self._xa.stack(data=(DIMS[1], DIMS[2])).values,
index=self._time_index,
columns=df_col_names,
)
else:
if copy:
return pd.DataFrame(
self._xa[:, :, 0].values.copy(),
index=self._time_index.copy(),
columns=self._xa.get_index(DIMS[1]).copy(),
)
else:
return pd.DataFrame(
self._xa[:, :, 0].values,
index=self._time_index,
columns=self._xa.get_index(DIMS[1]),
)
[docs] def quantile_df(self, quantile=0.5) -> pd.DataFrame:
"""
Return a Pandas DataFrame containing the single desired quantile of each component (over the samples).
Each of the series components will appear as a column in the DataFrame. The column will be named
"<component>_X", where "<component>" is the column name corresponding to this component, and "X"
is the quantile value.
The quantile columns represent the marginal distributions of the components of this series.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
quantile
The desired quantile value. The value must be represented as a fraction
(between 0 and 1 inclusive). For instance, `0.5` will return a DataFrame
containing the median of the (marginal) distribution of each component.
Returns
-------
pandas.DataFrame
The Pandas DataFrame containing the desired quantile for each component.
"""
self._assert_stochastic()
raise_if_not(
0 <= quantile <= 1,
"The quantile values must be expressed as fraction (between 0 and 1 inclusive).",
logger,
)
# column names
cnames = [s + f"_{quantile}" for s in self.columns]
return pd.DataFrame(
self._xa.quantile(q=quantile, dim=DIMS[2]),
index=self._time_index,
columns=cnames,
)
[docs] def quantile_timeseries(self, quantile=0.5, **kwargs) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the single desired quantile of each component
(over the samples) of this stochastic ``TimeSeries``.
The components in the new series are named "<component>_X", where "<component>"
is the column name corresponding to this component, and "X" is the quantile value.
The quantile columns represent the marginal distributions of the components of this series.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
quantile
The desired quantile value. The value must be represented as a fraction
(between 0 and 1 inclusive). For instance, `0.5` will return a TimeSeries
containing the median of the (marginal) distribution of each component.
kwargs
Other keyword arguments are passed down to `numpy.quantile()`
Returns
-------
TimeSeries
The TimeSeries containing the desired quantile for each component.
"""
self._assert_stochastic()
raise_if_not(
0 <= quantile <= 1,
"The quantile values must be expressed as fraction (between 0 and 1 inclusive).",
logger,
)
# component names
cnames = [f"{comp}_{quantile}" for comp in self.components]
new_data = np.quantile(
self._xa.values,
q=quantile,
axis=2,
overwrite_input=False,
keepdims=True,
**kwargs,
)
new_xa = xr.DataArray(
new_data,
dims=self._xa.dims,
coords={self._xa.dims[0]: self.time_index, DIMS[1]: pd.Index(cnames)},
attrs=self._xa.attrs,
)
return self.__class__(new_xa)
[docs] def quantiles_df(self, quantiles: Tuple[float] = (0.1, 0.5, 0.9)) -> pd.DataFrame:
"""
Return a Pandas DataFrame containing the desired quantiles of each component (over the samples).
Each of the series components will appear as a column in the DataFrame. The column will be named
"<component>_X", where "<component>" is the column name corresponding to this component, and "X"
is the quantile value.
The quantiles represent the marginal distributions of the components of this series.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
quantiles
Tuple containing the desired quantiles. The values must be represented as fractions
(between 0 and 1 inclusive). For instance, `(0.1, 0.5, 0.9)` will return a DataFrame
containing the 10th-percentile, median and 90th-percentile of the (marginal) distribution of each component.
Returns
-------
pandas.DataFrame
The Pandas DataFrame containing the quantiles for each component.
"""
return pd.concat(
[
self.quantile_timeseries(quantile).pd_dataframe()
for quantile in quantiles
],
axis=1,
)
[docs] def astype(self, dtype: Union[str, np.dtype]) -> Self:
"""
Converts this series to a new series with desired dtype.
Parameters
----------
dtype
A NumPy dtype (np.float32 or np.float64)
Returns
-------
TimeSeries
A TimeSeries having the desired dtype.
"""
return self.__class__(self._xa.astype(dtype))
[docs] def start_time(self) -> Union[pd.Timestamp, int]:
"""
Start time of the series.
Returns
-------
Union[pandas.Timestamp, int]
A timestamp containing the first time of the TimeSeries (if indexed by DatetimeIndex),
or an integer (if indexed by RangeIndex)
"""
return self._time_index[0]
[docs] def end_time(self) -> Union[pd.Timestamp, int]:
"""
End time of the series.
Returns
-------
Union[pandas.Timestamp, int]
A timestamp containing the last time of the TimeSeries (if indexed by DatetimeIndex),
or an integer (if indexed by RangeIndex)
"""
return self._time_index[-1]
[docs] def first_value(self) -> float:
"""
First value of this univariate series.
Returns
-------
float
The first value of this univariate deterministic time series
"""
self._assert_univariate()
self._assert_deterministic()
return float(self._xa[0, 0, 0])
[docs] def last_value(self) -> float:
"""
Last value of this univariate series.
Returns
-------
float
The last value of this univariate deterministic time series
"""
self._assert_univariate()
self._assert_deterministic()
return float(self._xa[-1, 0, 0])
[docs] def first_values(self) -> np.ndarray:
"""
First values of this potentially multivariate series.
Returns
-------
np.ndarray
The first values of every component of this deterministic time series
"""
self._assert_deterministic()
return self._xa.values[0, :, 0].copy()
[docs] def last_values(self) -> np.ndarray:
"""
Last values of this potentially multivariate series.
Returns
-------
np.ndarray
The last values of every component of this deterministic time series
"""
self._assert_deterministic()
return self._xa.values[-1, :, 0].copy()
[docs] def values(self, copy: bool = True, sample: int = 0) -> np.ndarray:
"""
Return a 2-D array of shape (time, component), containing this series' values for one `sample`.
Parameters
----------
copy
Whether to return a copy of the values, otherwise returns a view.
Leave it to True unless you know what you are doing.
sample
For stochastic series, the sample for which to return values. Default: 0 (first sample).
Returns
-------
numpy.ndarray
The values composing the time series.
"""
raise_if(
self.is_deterministic and sample != 0,
"This series contains one sample only (deterministic),"
"so only sample=0 is accepted.",
logger,
)
if copy:
return np.copy(self._xa.values[:, :, sample])
else:
return self._xa.values[:, :, sample]
[docs] def random_component_values(self, copy: bool = True) -> np.array:
"""
Return a 2-D array of shape (time, component), containing the values for
one sample taken uniformly at random among this series' samples.
Parameters
----------
copy
Whether to return a copy of the values, otherwise returns a view.
Leave it to True unless you know what you are doing.
Returns
-------
numpy.ndarray
The values composing one sample taken at random from the time series.
"""
sample = np.random.randint(low=0, high=self.n_samples)
if copy:
return np.copy(self._xa.values[:, :, sample])
else:
return self._xa.values[:, :, sample]
[docs] def all_values(self, copy: bool = True) -> np.ndarray:
"""
Return a 3-D array of dimension (time, component, sample),
containing this series' values for all samples.
Parameters
----------
copy
Whether to return a copy of the values, otherwise returns a view.
Leave it to True unless you know what you are doing.
Returns
-------
numpy.ndarray
The values composing the time series.
"""
if copy:
return np.copy(self._xa.values)
else:
return self._xa.values
[docs] def univariate_values(self, copy: bool = True, sample: int = 0) -> np.ndarray:
"""
Return a 1-D Numpy array of shape (time,),
containing this univariate series' values for one `sample`.
Parameters
----------
copy
Whether to return a copy of the values. Leave it to True unless you know what you are doing.
sample
For stochastic series, the sample for which to return values. Default: 0 (first sample).
Returns
-------
numpy.ndarray
The values composing the time series guaranteed to be univariate.
"""
self._assert_univariate()
if copy:
return np.copy(self._xa[:, 0, sample].values)
else:
return self._xa[:, 0, sample].values
[docs] def static_covariates_values(self, copy: bool = True) -> Optional[np.ndarray]:
"""
Return a 2-D array of dimension (component, static variable),
containing the static covariate values of the TimeSeries.
Parameters
----------
copy
Whether to return a copy of the values, otherwise returns a view.
Can only return a view if all values have the same dtype.
Leave it to True unless you know what you are doing.
Returns
-------
Optional[numpy.ndarray]
The static covariate values if the series has static covariates, else `None`.
"""
return (
self.static_covariates.to_numpy(copy=copy)
if self.has_static_covariates
else self.static_covariates
)
[docs] def head(
self, size: Optional[int] = 5, axis: Optional[Union[int, str]] = 0
) -> Self:
"""
Return a TimeSeries containing the first `size` points.
Parameters
----------
size : int, default 5
number of points to retain
axis : str or int, optional, default: 0
axis along which to slice the series
Returns
-------
TimeSeries
The series made of the first `size` points along the desired `axis`.
"""
axis_str = self._get_dim_name(axis)
display_n = min(size, self._xa.sizes[axis_str])
if axis_str == self._time_dim:
return self[:display_n]
else:
return self.__class__(self._xa[{axis_str: range(display_n)}])
[docs] def tail(
self, size: Optional[int] = 5, axis: Optional[Union[int, str]] = 0
) -> Self:
"""
Return last `size` points of the series.
Parameters
----------
size : int, default: 5
number of points to retain
axis : str or int, optional, default: 0 (time dimension)
axis along which we intend to display records
Returns
-------
TimeSeries
The series made of the last `size` points along the desired `axis`.
"""
axis_str = self._get_dim_name(axis)
display_n = min(size, self._xa.sizes[axis_str])
if axis_str == self._time_dim:
return self[-display_n:]
else:
return self.__class__(self._xa[{axis_str: range(-display_n, 0)}])
[docs] def concatenate(
self,
other: Self,
axis: Optional[Union[str, int]] = 0,
ignore_time_axis: Optional[bool] = False,
ignore_static_covariates: bool = False,
drop_hierarchy: bool = True,
) -> Self:
"""
Concatenate another timeseries to the current one along given axis.
Parameters
----------
other : TimeSeries
another timeseries to concatenate to this one
axis : str or int
axis along which timeseries will be concatenated. ['time', 'component' or 'sample'; Default: 0 (time)]
ignore_time_axis : bool, default False
Ignore errors when time axis varies for some timeseries. Note that this may yield unexpected results
ignore_static_covariates : bool
whether to ignore all requirements for static covariate concatenation and only transfer the
static covariates of the first TimeSeries element in `series` to the concatenated TimeSeries.
Only effective when `axis=1`.
drop_hierarchy : bool
When `axis=1`, whether to drop hierarchy information. True by default.
When False, the hierarchies will be "concatenated" as well
(by merging the hierarchy dictionaries), which may cause issues if the component
names of the resulting series and that of the merged hierarchy do not match.
When `axis=0` or `axis=2`, the hierarchy of the first series is always kept.
Returns
-------
TimeSeries
concatenated timeseries
See Also
--------
concatenate : a function to concatenate multiple series along a given axis.
Notes
-----
When concatenating along the `time` dimension, the current series marks the start date of
the resulting series, and the other series will have its time index ignored.
"""
return concatenate(
series=[self, other],
axis=axis,
ignore_time_axis=ignore_time_axis,
ignore_static_covariates=ignore_static_covariates,
drop_hierarchy=drop_hierarchy,
)
"""
Other methods
=============
"""
[docs] def gaps(self, mode: Literal["all", "any"] = "all") -> pd.DataFrame:
"""
A function to compute and return gaps in the TimeSeries. Works only on deterministic time series (1 sample).
Parameters
----------
mode
Only relevant for multivariate time series. The mode defines how gaps are defined. Set to
'any' if a NaN value in any columns should be considered as as gaps. 'all' will only
consider periods where all columns' values are NaN. Defaults to 'all'.
Returns
-------
pd.DataFrame
A pandas.DataFrame containing a row for every gap (rows with all-NaN values in underlying DataFrame)
in this time series. The DataFrame contains three columns that include the start and end time stamps
of the gap and the integer length of the gap (in `self.freq` units if the series is indexed
by a DatetimeIndex).
"""
df = self.pd_dataframe()
if mode == "all":
is_nan_series = df.isna().all(axis=1).astype(int)
elif mode == "any":
is_nan_series = df.isna().any(axis=1).astype(int)
else:
raise_log(
ValueError(
f"Keyword mode accepts only 'any' or 'all'. Provided {mode}"
),
logger,
)
diff = pd.Series(np.diff(is_nan_series.values), index=is_nan_series.index[:-1])
gap_starts = diff[diff == 1].index + self._freq
gap_ends = diff[diff == -1].index
if is_nan_series.iloc[0] == 1:
gap_starts = gap_starts.insert(0, self.start_time())
if is_nan_series.iloc[-1] == 1:
gap_ends = gap_ends.insert(len(gap_ends), self.end_time())
gap_df = pd.DataFrame(columns=["gap_start", "gap_end"])
if gap_starts.size == 0:
return gap_df
else:
def intvl(start, end):
if self._has_datetime_index:
return pd.date_range(start=start, end=end, freq=self._freq).size
else:
return int((end - start) / self._freq) + 1
gap_df["gap_start"] = gap_starts
gap_df["gap_end"] = gap_ends
gap_df["gap_size"] = gap_df.apply(
lambda row: intvl(start=row.gap_start, end=row.gap_end), axis=1
)
return gap_df
[docs] def copy(self) -> Self:
"""
Make a copy of this series.
Returns
-------
TimeSeries
A copy of this time series.
"""
# the xarray will be copied in the TimeSeries constructor.
return self.__class__(self._xa)
[docs] def get_index_at_point(
self, point: Union[pd.Timestamp, float, int], after=True
) -> int:
"""
Converts a point along the time axis index into an integer index ranging in (0, len(series)-1).
Parameters
----------
point
This parameter supports 3 different data types: ``pd.Timestamp``, ``float`` and ``int``.
``pd.Timestamp`` work only on series that are indexed with a ``pd.DatetimeIndex``. In such cases, the
returned point will be the index of this timestamp if it is present in the series time index.
It it's not present in the time index, the index of the next timestamp is returned if `after=True`
(if it exists in the series), otherwise the index of the previous timestamp is returned
(if it exists in the series).
In case of a ``float``, the parameter will be treated as the proportion of the time series
that should lie before the point.
If an ``int`` and series is datetime-indexed, the value of `point` is returned.
If an ``int`` and series is integer-indexed, the index position of `point` in the RangeIndex is returned
(accounting for steps).
after
If the provided pandas Timestamp is not in the time series index, whether to return the index of the
next timestamp or the index of the previous one.
"""
point_index = -1
if isinstance(point, float):
raise_if_not(
0.0 <= point <= 1.0,
"point (float) should be between 0.0 and 1.0.",
logger,
)
point_index = int((len(self) - 1) * point)
elif isinstance(point, (int, np.int64)):
if self.has_datetime_index or (self.start_time() == 0 and self.freq == 1):
point_index = point
else:
point_index_float = (point - self.start_time()) / self.freq
point_index = int(point_index_float)
raise_if(
point_index != point_index_float,
"The provided point is not a valid index for this series.",
)
raise_if_not(
0 <= point_index < len(self),
f"The index corresponding to the provided point ({point}) should be a valid index in series",
logger,
)
elif isinstance(point, pd.Timestamp):
raise_if_not(
self._has_datetime_index,
"A Timestamp has been provided, but this series is not time-indexed.",
logger,
)
self._raise_if_not_within(point)
if point in self:
point_index = self._time_index.get_loc(point)
else:
point_index = self._time_index.get_loc(
self._get_first_timestamp_after(point)
if after
else self._get_last_timestamp_before(point)
)
else:
raise_log(
TypeError(
"`point` needs to be either `float`, `int` or `pd.Timestamp`"
),
logger,
)
return point_index
[docs] def get_timestamp_at_point(
self, point: Union[pd.Timestamp, float, int]
) -> Union[pd.Timestamp, int]:
"""
Converts a point into a pandas.Timestamp (if Datetime-indexed) or into an integer (if Int64-indexed).
Parameters
----------
point
This parameter supports 3 different data types: `float`, `int` and `pandas.Timestamp`.
In case of a `float`, the parameter will be treated as the proportion of the time series
that should lie before the point.
In case of `int`, the parameter will be treated as an integer index to the time index of
`series`. Will raise a ValueError if not a valid index in `series`.
In case of a `pandas.Timestamp`, point will be returned as is provided that the timestamp
is present in the series time index, otherwise will raise a ValueError.
"""
idx = self.get_index_at_point(point)
return self._time_index[idx]
def _split_at(
self, split_point: Union[pd.Timestamp, float, int], after: bool = True
) -> Tuple[Self, Self]:
# Get index with not after in order to avoid moving twice if split_point is not in self
point_index = self.get_index_at_point(split_point, not after)
return (
self[: point_index + (1 if after else 0)],
self[point_index + (1 if after else 0) :],
)
[docs] def split_after(
self, split_point: Union[pd.Timestamp, float, int]
) -> Tuple[Self, Self]:
"""
Splits the series in two, after a provided `split_point`.
Parameters
----------
split_point
A timestamp, float or integer. If float, represents the proportion of the series to include in the
first TimeSeries (must be between 0.0 and 1.0). If integer, represents the index position after
which the split is performed. A pd.Timestamp can be provided for TimeSeries that are indexed by a
pd.DatetimeIndex. In such cases, the timestamp will be contained in the first TimeSeries, but not
in the second one. The timestamp itself does not have to appear in the original TimeSeries index.
Returns
-------
Tuple[TimeSeries, TimeSeries]
A tuple of two time series. The first time series contains the first samples up to the `split_point`,
and the second contains the remaining ones.
"""
return self._split_at(split_point, after=True)
[docs] def split_before(
self, split_point: Union[pd.Timestamp, float, int]
) -> Tuple[Self, Self]:
"""
Splits the series in two, before a provided `split_point`.
Parameters
----------
split_point
A timestamp, float or integer. If float, represents the proportion of the series to include in the
first TimeSeries (must be between 0.0 and 1.0). If integer, represents the index position before
which the split is performed. A pd.Timestamp can be provided for TimeSeries that are indexed by a
pd.DatetimeIndex. In such cases, the timestamp will be contained in the second TimeSeries, but not
in the first one. The timestamp itself does not have to appear in the original TimeSeries index.
Returns
-------
Tuple[TimeSeries, TimeSeries]
A tuple of two time series. The first time series contains the first samples up to the `split_point`,
and the second contains the remaining ones.
"""
return self._split_at(split_point, after=False)
[docs] def drop_after(self, split_point: Union[pd.Timestamp, float, int]):
"""
Drops everything after the provided time `split_point`, included.
The timestamp may not be in the series. If it is, the timestamp will be dropped.
Parameters
----------
split_point
The timestamp that indicates cut-off time.
Returns
-------
TimeSeries
A new TimeSeries, after `ts`.
"""
return self[: self.get_index_at_point(split_point, after=True)]
[docs] def drop_before(self, split_point: Union[pd.Timestamp, float, int]):
"""
Drops everything before the provided time `split_point`, included.
The timestamp may not be in the series. If it is, the timestamp will be dropped.
Parameters
----------
split_point
The timestamp that indicates cut-off time.
Returns
-------
TimeSeries
A new TimeSeries, after `ts`.
"""
return self[self.get_index_at_point(split_point, after=False) + 1 :]
[docs] def slice(
self, start_ts: Union[pd.Timestamp, int], end_ts: Union[pd.Timestamp, int]
):
"""
Return a new TimeSeries, starting later than `start_ts` and ending before `end_ts`.
For series having DatetimeIndex, this is inclusive on both ends. For series having a RangeIndex,
`end_ts` is exclusive.
`start_ts` and `end_ts` don't have to be in the series.
Parameters
----------
start_ts
The timestamp that indicates the left cut-off.
end_ts
The timestamp that indicates the right cut-off.
Returns
-------
TimeSeries
A new series, with indices greater or equal than `start_ts` and smaller or equal than `end_ts`.
"""
raise_if_not(
type(start_ts) is type(end_ts),
"The two timestamps provided to slice() have to be of the same type.",
logger,
)
if isinstance(start_ts, pd.Timestamp):
raise_if_not(
self._has_datetime_index,
"Timestamps have been provided to slice(), but the series is "
"indexed using an integer-based RangeIndex.",
logger,
)
if start_ts in self._time_index and end_ts in self._time_index:
return self[
start_ts:end_ts
] # we assume this is faster than the filtering below
else:
idx = self._time_index[
(start_ts <= self._time_index) & (self._time_index <= end_ts)
]
return self[idx]
else:
raise_if(
self._has_datetime_index,
"start and end times have been provided as integers to slice(), but "
"the series is indexed with a DatetimeIndex.",
logger,
)
# get closest timestamps if either start or end are not in the index
effective_start_ts = (
min(self._time_index, key=lambda t: abs(t - start_ts))
if start_ts not in self._time_index
else start_ts
)
if effective_start_ts < start_ts:
# if the requested start_ts is smaller than the start argument,
# we have to increase it to be consistent with the docstring
effective_start_ts += self.freq
effective_end_ts = (
min(self._time_index, key=lambda t: abs(t - end_ts))
if end_ts not in self._time_index
else end_ts
)
if end_ts >= effective_end_ts + self.freq:
# if the requested end_ts is further off from the end of the time series,
# we have to increase effectiv_end_ts to make the last timestamp inclusive.
effective_end_ts += self.freq
idx = pd.RangeIndex(effective_start_ts, effective_end_ts, step=self.freq)
return self[idx]
[docs] def slice_n_points_after(self, start_ts: Union[pd.Timestamp, int], n: int) -> Self:
"""
Return a new TimeSeries, starting a `start_ts` (inclusive) and having at most `n` points.
The provided timestamps will be included in the series.
Parameters
----------
start_ts
The timestamp or index that indicates the splitting time.
n
The maximal length of the new TimeSeries.
Returns
-------
TimeSeries
A new TimeSeries, with length at most `n`, starting at `start_ts`
"""
raise_if_not(n > 0, "n should be a positive integer.", logger)
self._raise_if_not_within(start_ts)
if isinstance(start_ts, (int, np.int64)):
return self[pd.RangeIndex(start=start_ts, stop=start_ts + n)]
elif isinstance(start_ts, pd.Timestamp):
# get first timestamp greater or equal to start_ts
tss = self._get_first_timestamp_after(start_ts)
point_index = self.get_index_at_point(tss)
return self[point_index : point_index + n]
else:
raise_log(
ValueError("start_ts must be an int or a pandas Timestamp."), logger
)
[docs] def slice_n_points_before(self, end_ts: Union[pd.Timestamp, int], n: int) -> Self:
"""
Return a new TimeSeries, ending at `end_ts` (inclusive) and having at most `n` points.
The provided timestamps will be included in the series.
Parameters
----------
end_ts
The timestamp or index that indicates the splitting time.
n
The maximal length of the new TimeSeries.
Returns
-------
TimeSeries
A new TimeSeries, with length at most `n`, ending at `start_ts`
"""
raise_if_not(n > 0, "n should be a positive integer.", logger)
self._raise_if_not_within(end_ts)
if isinstance(end_ts, (int, np.int64)):
return self[pd.RangeIndex(start=end_ts - n + 1, stop=end_ts + 1)]
elif isinstance(end_ts, pd.Timestamp):
# get last timestamp smaller or equal to start_ts
tss = self._get_last_timestamp_before(end_ts)
point_index = self.get_index_at_point(tss)
return self[max(0, point_index - n + 1) : point_index + 1]
else:
raise_log(
ValueError("start_ts must be an int or a pandas Timestamp."), logger
)
[docs] def slice_intersect(self, other: Self) -> Self:
"""
Return a ``TimeSeries`` slice of this series, where the time index has been intersected with the one
of the `other` series.
This method is in general *not* symmetric.
Parameters
----------
other
the other time series
Returns
-------
TimeSeries
a new series, containing the values of this series, over the time-span common to both time series.
"""
if other.has_same_time_as(self):
return self.__class__(self._xa)
if other.freq == self.freq:
start, end = self._slice_intersect_bounds(other)
return self[start:end]
else:
time_index = self.time_index.intersection(other.time_index)
return self[time_index]
[docs] def slice_intersect_values(self, other: Self, copy: bool = False) -> Self:
"""
Return the sliced values of this series, where the time index has been intersected with the one
of the `other` series.
This method is in general *not* symmetric.
Parameters
----------
other
The other time series
copy
Whether to return a copy of the values, otherwise returns a view.
Leave it to True unless you know what you are doing.
Returns
-------
np.ndarray
The values of this series, over the time-span common to both time series.
"""
vals = self.all_values(copy=copy)
if other.has_same_time_as(self):
return vals
if other.freq == self.freq:
start, end = self._slice_intersect_bounds(other)
return vals[start:end]
else:
return vals[self.time_index.isin(other.time_index)]
def _slice_intersect_bounds(self, other: Self) -> Tuple[int, int]:
shift_start = n_steps_between(
other.start_time(), self.start_time(), freq=self.freq
)
shift_end = n_steps_between(other.end_time(), self.end_time(), freq=self.freq)
shift_start = shift_start if shift_start >= 0 else 0
shift_end = shift_end if shift_end < 0 else None
return shift_start, shift_end
[docs] def strip(self, how: str = "all") -> Self:
"""
Return a ``TimeSeries`` slice of this deterministic time series, where NaN-containing entries at the beginning
and the end of the series are removed. No entries after (and including) the first non-NaN entry and
before (and including) the last non-NaN entry are removed.
This method is only applicable to deterministic series (i.e., having 1 sample).
Parameters
----------
how
Define if the entries containing `NaN` in all the components ('all') or in any of the components ('any')
should be stripped. Default: 'all'
Returns
-------
TimeSeries
a new series based on the original where NaN-containing entries at start and end have been removed
"""
raise_if(
self.is_probabilistic,
"`strip` cannot be applied to stochastic TimeSeries",
logger,
)
first_finite_row, last_finite_row = _finite_rows_boundaries(
self.values(), how=how
)
return self.__class__.from_times_and_values(
times=self.time_index[first_finite_row : last_finite_row + 1],
values=self.values()[first_finite_row : last_finite_row + 1],
columns=self.components,
static_covariates=self.static_covariates,
hierarchy=self.hierarchy,
)
[docs] def longest_contiguous_slice(
self, max_gap_size: int = 0, mode: str = "all"
) -> Self:
"""
Return the largest TimeSeries slice of this deterministic series that contains no gaps
(contiguous all-NaN values) larger than `max_gap_size`.
This method is only applicable to deterministic series (i.e., having 1 sample).
Parameters
----------
max_gap_size
Indicate the maximum gap size that the TimeSerie can contain
mode
Only relevant for multivariate time series. The mode defines how gaps are defined. Set to
'any' if a NaN value in any columns should be considered as as gaps. 'all' will only
consider periods where all columns' values are NaN. Defaults to 'all'.
Returns
-------
TimeSeries
a new series constituting the largest slice of the original with no or bounded gaps
See Also
--------
TimeSeries.gaps : return the gaps in the TimeSeries
"""
if not (np.isnan(self._xa)).any():
return self.copy()
stripped_series = self.strip()
gaps = stripped_series.gaps(mode=mode)
relevant_gaps = gaps[gaps["gap_size"] > max_gap_size]
curr_slice_start = stripped_series.start_time()
max_size = pd.Timedelta(days=0) if self._has_datetime_index else 0
max_slice_start = None
max_slice_end = None
for index, row in relevant_gaps.iterrows():
# evaluate size of the current slice. the slice ends one time step before row['gap_start']
curr_slice_end = row["gap_start"] - self.freq
size = curr_slice_end - curr_slice_start
if size > max_size:
max_size = size
max_slice_start = curr_slice_start
max_slice_end = row["gap_start"] - self._freq
curr_slice_start = row["gap_end"] + self._freq
if stripped_series.end_time() - curr_slice_start > max_size:
max_slice_start = curr_slice_start
max_slice_end = self.end_time()
return stripped_series[max_slice_start:max_slice_end]
[docs] def rescale_with_value(self, value_at_first_step: float) -> Self:
"""
Return a new ``TimeSeries``, which is a multiple of this series such that
the first value is `value_at_first_step`.
(Note: numerical errors can appear with `value_at_first_step > 1e+24`).
Parameters
----------
value_at_first_step
The new value for the first entry of the TimeSeries.
Returns
-------
TimeSeries
A new TimeSeries, where the first value is `value_at_first_step` and other values
have been scaled accordingly.
"""
raise_if_not(
(self._xa[0, :, :] != 0).all(), "Cannot rescale with first value 0.", logger
)
coef = value_at_first_step / self._xa.isel({self._time_dim: [0]})
coef = coef.values.reshape((self.n_components, self.n_samples)) # TODO: test
new_series = coef * self._xa
return self.__class__(new_series)
[docs] def shift(self, n: int) -> Self:
"""
Shifts the time axis of this TimeSeries by `n` time steps.
If :math:`n > 0`, shifts in the future. If :math:`n < 0`, shifts in the past.
For example, with :math:`n=2` and `freq='M'`, March 2013 becomes May 2013.
With :math:`n=-2`, March 2013 becomes Jan 2013.
Parameters
----------
n
The number of time steps (in self.freq unit) to shift by. Can be negative.
Returns
-------
TimeSeries
A new TimeSeries, with a shifted index.
"""
if not isinstance(n, (int, np.int64)):
logger.warning(
f"TimeSeries.shift(): converting n to int from {n} to {int(n)}"
)
n = int(n)
try:
self._time_index[-1] + n * self.freq
except pd.errors.OutOfBoundsDatetime:
raise_log(
OverflowError(
"the add operation between {} and {} will "
"overflow".format(n * self.freq, self.time_index[-1])
),
logger,
)
if self.has_range_index:
new_time_index = self._time_index + n * self.freq
else:
new_time_index = self._time_index.map(lambda ts: ts + n * self.freq)
if new_time_index.freq is None:
new_time_index.freq = self.freq
new_xa = self._xa.assign_coords({self._xa.dims[0]: new_time_index})
return self.__class__(new_xa)
[docs] def diff(
self,
n: Optional[int] = 1,
periods: Optional[int] = 1,
dropna: Optional[bool] = True,
) -> Self:
"""
Return a differenced time series. This is often used to make a time series stationary.
Parameters
----------
n
Optionally, a positive integer indicating the number of differencing steps (default = 1).
For instance, n=2 computes the second order differences.
periods
Optionally, periods to shift for calculating difference. For instance, periods=12 computes the
difference between values at time `t` and times `t-12`.
dropna
Whether to drop the missing values after each differencing steps. If set to False, the corresponding
first `periods` time steps will be filled with NaNs.
Returns
-------
TimeSeries
A new TimeSeries, with the differenced values.
"""
if not isinstance(n, int) or n < 1:
raise_log(ValueError("'n' must be a positive integer >= 1."), logger)
if not isinstance(periods, int) or periods < 1:
raise_log(ValueError("'periods' must be an integer >= 1."), logger)
def _compute_diff(xa: xr.DataArray):
# xarray doesn't support Pandas "period" so compute diff() ourselves
if not dropna:
# In this case the new DataArray will have the same size and filled with NaNs
new_xa_ = xa.copy()
new_xa_.values[:periods, :, :] = np.nan
new_xa_.values[periods:, :, :] = (
xa.values[periods:, :, :] - xa.values[:-periods, :, :]
)
else:
# In this case the new DataArray will be shorter
new_xa_ = xa[periods:, :, :].copy()
new_xa_.values = xa.values[periods:, :, :] - xa.values[:-periods, :, :]
return new_xa_
new_xa = _compute_diff(self._xa)
for _ in range(n - 1):
new_xa = _compute_diff(new_xa)
return self.__class__(new_xa)
[docs] def cumsum(self) -> Self:
"""
Returns the cumulative sum of the time series along the time axis.
Returns
-------
TimeSeries
A new TimeSeries, with the cumulatively summed values.
"""
return self.__class__(self._xa.copy().cumsum(axis=0))
[docs] def has_same_time_as(self, other: Self) -> bool:
"""
Checks whether this series has the same time index as `other`.
Parameters
----------
other
the other series
Returns
-------
bool
True if both TimeSeries have the same index, False otherwise.
"""
if len(other) != len(self):
return False
if other.freq != self.freq:
return False
if other.start_time() != self.start_time():
return False
else:
return True
[docs] def append(self, other: Self) -> Self:
"""
Appends another series to this series along the time axis.
Parameters
----------
other
A second TimeSeries.
Returns
-------
TimeSeries
A new TimeSeries, obtained by appending the second TimeSeries to the first.
See Also
--------
TimeSeries.concatenate : concatenate another series along a given axis.
TimeSeries.prepend : prepend (i.e. add to the beginning) another series along the time axis.
"""
raise_if_not(
other.has_datetime_index == self.has_datetime_index,
"Both series must have the same type of time index (either DatetimeIndex or RangeIndex).",
logger,
)
raise_if_not(
other.freq == self.freq,
"Both series must have the same frequency.",
logger,
)
raise_if_not(
other.n_components == self.n_components,
"Both series must have the same number of components.",
logger,
)
raise_if_not(
other.n_samples == self.n_samples,
"Both series must have the same number of components.",
logger,
)
if self._has_datetime_index:
raise_if_not(
other.start_time() == self.end_time() + self.freq,
"Appended TimeSeries must start one time step after current one.",
logger,
)
other_xa = other.data_array()
new_xa = xr.DataArray(
np.concatenate((self._xa.values, other_xa.values), axis=0),
dims=self._xa.dims,
coords={
self._time_dim: self._time_index.append(other.time_index),
DIMS[1]: self.components,
},
attrs=self._xa.attrs,
)
return self.__class__.from_xarray(
new_xa, fill_missing_dates=True, freq=self._freq_str
)
[docs] def append_values(self, values: np.ndarray) -> Self:
"""
Appends new values to current TimeSeries, extending its time index.
Parameters
----------
values
An array with the values to append.
Returns
-------
TimeSeries
A new TimeSeries with the new values appended
"""
if self._has_datetime_index:
idx = pd.DatetimeIndex(
[self.end_time() + i * self._freq for i in range(1, len(values) + 1)],
freq=self._freq,
)
else:
idx = pd.RangeIndex(
start=self.end_time() + self._freq,
stop=self.end_time() + (len(values) + 1) * self._freq,
step=self._freq,
)
return self.append(
self.__class__.from_times_and_values(
values=values,
times=idx,
fill_missing_dates=False,
static_covariates=self.static_covariates,
)
)
[docs] def prepend(self, other: Self) -> Self:
"""
Prepends (i.e. adds to the beginning) another series to this series along the time axis.
Parameters
----------
other
A second TimeSeries.
Returns
-------
TimeSeries
A new TimeSeries, obtained by appending the second TimeSeries to the first.
See Also
--------
Timeseries.append : append (i.e. add to the end) another series along the time axis.
TimeSeries.concatenate : concatenate another series along a given axis.
"""
raise_if_not(
isinstance(other, self.__class__),
f"`other` to prepend must be a {self.__class__.__name__} object.",
)
return other.append(self)
[docs] def prepend_values(self, values: np.ndarray) -> Self:
"""
Prepends (i.e. adds to the beginning) new values to current TimeSeries, extending its time index into the past.
Parameters
----------
values
An array with the values to prepend to the start.
Returns
-------
TimeSeries
A new TimeSeries with the new values prepended.
"""
if self._has_datetime_index:
idx = pd.DatetimeIndex(
[
self.start_time() - i * self._freq
for i in reversed(range(1, len(values) + 1))
],
freq=self._freq,
)
else:
idx = pd.RangeIndex(
self.start_time() - self.freq * len(values),
self.start_time(),
step=self.freq,
)
return self.prepend(
self.__class__.from_times_and_values(
values=values,
times=idx,
fill_missing_dates=False,
static_covariates=self.static_covariates,
columns=self.columns,
hierarchy=self.hierarchy,
)
)
[docs] def with_values(self, values: np.ndarray) -> Self:
"""
Return a new ``TimeSeries`` similar to this one but with new specified values.
Parameters
----------
values
A Numpy array with new values. It must have the dimensions for time
and componentns, but may contain a different number of samples.
Returns
-------
TimeSeries
A new TimeSeries with the new values and same index, static covariates and hierarchy
"""
raise_if_not(
values.shape[:2] == self._xa.values.shape[:2],
"The new values must have the same shape (time, components) as the present series. "
"Received: {}, expected: {}".format(
values.shape[:2], self._xa.values.shape[:2]
),
)
new_xa = xr.DataArray(
values,
dims=self._xa.dims,
coords=self._xa.coords,
attrs=self._xa.attrs,
)
return self.__class__(new_xa)
[docs] def with_static_covariates(
self, covariates: Optional[Union[pd.Series, pd.DataFrame]]
):
"""Returns a new TimeSeries object with added static covariates.
Static covariates contain data attached to the time series, but which are not varying with time.
Parameters
----------
covariates
Optionally, a set of static covariates to be added to the TimeSeries. Either a pandas Series, a pandas
DataFrame, or `None`. If `None`, will set the static covariates to `None`. If a Series, the index
represents the static variables. The covariates are then globally 'applied' to all components of the
TimeSeries. If a DataFrame, the columns represent the static variables and the rows represent the
components of the uni/multivariate TimeSeries. If a single-row DataFrame, the covariates are globally
'applied' to all components of the TimeSeries. If a multi-row DataFrame, the number of rows must match the
number of components of the TimeSeries. This adds component-specific static covariates.
Notes
-----
If there are a large number of static covariates variables (i.e., the static covariates have a very large
dimension), there might be a noticeable performance penalty for creating ``TimeSeries`` objects, unless
the covariates already have the same ``dtype`` as the series data.
Examples
--------
>>> import pandas as pd
>>> from darts.utils.timeseries_generation import linear_timeseries
>>> # add global static covariates
>>> static_covs = pd.Series([0., 1.], index=["static_cov_1", "static_cov_2"])
>>> series = linear_timeseries(length=3)
>>> series_new1 = series.with_static_covariates(static_covs)
>>> series_new1.static_covariates
static_cov_1 static_cov_2
component
linear 0.0 1.0
>>> # add component specific static covariates
>>> static_covs_multi = pd.DataFrame([[0., 1.], [2., 3.]], columns=["static_cov_1", "static_cov_2"])
>>> series_multi = series.stack(series)
>>> series_new2 = series_multi.with_static_covariates(static_covs_multi)
>>> series_new2.static_covariates
static_cov_1 static_cov_2
component
linear 0.0 1.0
linear_1 2.0 3.0
"""
return self.__class__(
xr.DataArray(
self._xa.values,
dims=self._xa.dims,
coords=self._xa.coords,
attrs={
STATIC_COV_TAG: covariates,
HIERARCHY_TAG: self.hierarchy,
},
)
)
[docs] def with_hierarchy(self, hierarchy: Dict[str, Union[str, List[str]]]):
"""
Adds a hierarchy to the TimeSeries.
Parameters
----------
hierarchy
A dictionary mapping components to a list of their parent(s) in the hierarchy.
Single parents may be specified as string or list containing one string.
For example, assume the series contains the components
``["total", "a", "b", "x", "y", "ax", "ay", "bx", "by"]``,
the following dictionary would encode the groupings shown on
`this figure <https://otexts.com/fpp3/hts.html#fig:GroupTree>`_:
.. highlight:: python
.. code-block:: python
hierarchy = {'ax': ['a', 'x'],
'ay': ['a', 'y'],
'bx': ['b', 'x'],
'by': ['b', 'y'],
'a': ['total'],
'b': ['total'],
'x': 'total', # or use a single string
'y': 'total'}
..
"""
return self.__class__(
xr.DataArray(
self._xa.values,
dims=self._xa.dims,
coords=self._xa.coords,
attrs={
STATIC_COV_TAG: self.static_covariates,
HIERARCHY_TAG: hierarchy,
},
)
)
[docs] def stack(self, other: Self) -> Self:
"""
Stacks another univariate or multivariate TimeSeries with the same time index on top of
the current one (along the component axis).
Return a new TimeSeries that includes all the components of `self` and of `other`.
The resulting TimeSeries will have the same name for its time dimension as this TimeSeries, and the
same number of samples.
Parameters
----------
other
A TimeSeries instance with the same index and the same number of samples as the current one.
Returns
-------
TimeSeries
A new multivariate TimeSeries instance.
"""
return concatenate([self, other], axis=1)
[docs] def drop_columns(self, col_names: Union[List[str], str]) -> Self:
"""
Return a new ``TimeSeries`` instance with dropped columns/components.
Parameters
-------
col_names
String or list of strings corresponding to the columns to be dropped.
Returns
-------
TimeSeries
A new TimeSeries instance with specified columns dropped.
"""
if isinstance(col_names, str):
col_names = [col_names]
raise_if_not(
all([(x in self.columns.to_list()) for x in col_names]),
"Some column names in col_names don't exist in the time series.",
logger,
)
new_xa = self._xa.drop_sel({"component": col_names})
return self.__class__(new_xa)
[docs] def univariate_component(self, index: Union[str, int]) -> Self:
"""
Retrieve one of the components of the series
and return it as new univariate ``TimeSeries`` instance.
This drops the hierarchy (if any), and retains only the relevant static
covariates column.
Parameters
----------
index
An zero-indexed integer indicating which component to retrieve. If components have names,
this can be a string with the component's name.
Returns
-------
TimeSeries
A new univariate TimeSeries instance.
"""
return self[index if isinstance(index, str) else self.components[index]]
[docs] def add_datetime_attribute(
self,
attribute,
one_hot: bool = False,
cyclic: bool = False,
tz: Optional[str] = None,
) -> Self:
"""
Build a new series with one (or more) additional component(s) that contain an attribute
of the time index of the series.
The additional components are specified with `attribute`, such as 'weekday', 'day' or 'month'.
This works only for deterministic time series (i.e., made of 1 sample).
Parameters
----------
attribute
A pd.DatatimeIndex attribute which will serve as the basis of the new column(s).
one_hot
Boolean value indicating whether to add the specified attribute as a one hot encoding
(results in more columns).
cyclic
Boolean value indicating whether to add the specified attribute as a cyclic encoding.
Alternative to one_hot encoding, enable only one of the two.
(adds 2 columns, corresponding to sin and cos transformation).
tz
Optionally, a time zone to convert the time index to before computing the attributes.
Returns
-------
TimeSeries
New TimeSeries instance enhanced by `attribute`.
"""
self._assert_deterministic()
from .utils import timeseries_generation as tg
return self.stack(
tg.datetime_attribute_timeseries(
self.time_index,
attribute=attribute,
one_hot=one_hot,
cyclic=cyclic,
tz=tz,
)
)
[docs] def add_holidays(
self,
country_code: str,
prov: str = None,
state: str = None,
tz: Optional[str] = None,
) -> Self:
"""
Adds a binary univariate component to the current series that equals 1 at every index that
corresponds to selected country's holiday, and 0 otherwise.
The frequency of the TimeSeries is daily.
Available countries can be found `here <https://github.com/dr-prodigy/python-holidays#available-countries>`_.
This works only for deterministic time series (i.e., made of 1 sample).
Parameters
----------
country_code
The country ISO code
prov
The province
state
The state
tz
Optionally, a time zone to convert the time index to before computing the attributes.
Returns
-------
TimeSeries
A new TimeSeries instance, enhanced with binary holiday component.
"""
self._assert_deterministic()
from .utils import timeseries_generation as tg
return self.stack(
tg.holidays_timeseries(
self.time_index,
country_code=country_code,
prov=prov,
state=state,
tz=tz,
)
)
[docs] def resample(self, freq: str, method: str = "pad", **kwargs) -> Self:
"""
Build a reindexed ``TimeSeries`` with a given frequency.
Provided method is used to fill holes in reindexed TimeSeries, by default 'pad'.
Parameters
----------
freq
The new time difference between two adjacent entries in the returned TimeSeries.
A DateOffset alias is expected.
method:
Method to fill holes in reindexed TimeSeries (note this does not fill NaNs that already were present):
'pad': propagate last valid observation forward to next valid
'backfill': use NEXT valid observation to fill.
kwargs
some keyword arguments for the `xarray.resample` method, notably `offset` or `base` to indicate where
to start the resampling and avoid nan at the first value of the resampled TimeSeries
For more informations, see the `xarray resample() documentation
<https://docs.xarray.dev/en/stable/generated/xarray.DataArray.resample.html>`_.
Examples
--------
>>> times = pd.date_range(start=pd.Timestamp("20200101233000"), periods=6, freq="15T")
>>> pd_series = pd.Series(range(6), index=times)
>>> ts = TimeSeries.from_series(pd_series)
>>> print(ts.time_index)
DatetimeIndex(['2020-01-01 23:30:00', '2020-01-01 23:45:00',
'2020-01-02 00:00:00', '2020-01-02 00:15:00',
'2020-01-02 00:30:00', '2020-01-02 00:45:00'],
dtype='datetime64[ns]', name='time', freq='15T')
>>> resampled_nokwargs_ts = ts.resample(freq="1h")
>>> print(resampled_nokwargs_ts.time_index)
DatetimeIndex(['2020-01-01 23:00:00', '2020-01-02 00:00:00'],
dtype='datetime64[ns]', name='time', freq='H')
>>> print(resampled_nokwargs_ts.values())
[[nan]
[ 2.]]
>>> resampled_ts = ts.resample(freq="1h", offset=pd.Timedelta("30T"))
>>> print(resampled_ts.time_index)
DatetimeIndex(['2020-01-01 23:30:00', '2020-01-02 00:30:00'],
dtype='datetime64[ns]', name='time', freq='H')
>>> print(resampled_ts.values())
[[0.]
[4.]]
Returns
-------
TimeSeries
A reindexed TimeSeries with given frequency.
"""
resample = self._xa.resample(
indexer={self._time_dim: freq},
**kwargs,
)
# TODO: check
if method == "pad":
new_xa = resample.pad()
elif method in ["bfill", "backfill"]:
new_xa = resample.backfill()
else:
raise_log(ValueError(f"Unknown method: {method}"), logger)
return self.__class__(new_xa)
[docs] def is_within_range(self, ts: Union[pd.Timestamp, int]) -> bool:
"""
Check whether a given timestamp or integer is within the time interval of this time series.
If a timestamp is provided, it does not need to be an element of the time index of the series.
Parameters
----------
ts
The `pandas.Timestamp` (if indexed with DatetimeIndex) or integer (if indexed with RangeIndex) to check.
Returns
-------
bool
Whether `ts` is contained within the interval of this time series.
"""
return self.time_index[0] <= ts <= self.time_index[-1]
[docs] def map(
self,
fn: Union[
Callable[[np.number], np.number],
Callable[[Union[pd.Timestamp, int], np.number], np.number],
],
) -> Self: # noqa: E501
"""
Applies the function `fn` to the underlying NumPy array containing this series' values.
Return a new TimeSeries instance. If `fn` takes 1 argument it is simply applied on the backing array
of shape (time, n_components, n_samples).
If it takes 2 arguments, it is applied repeatedly on the (ts, value[ts]) tuples, where
"ts" denotes a timestamp value, and "value[ts]" denote the array of values at this timestamp, of shape
(n_components, n_samples).
Parameters
----------
fn
Either a function which takes a NumPy array and returns a NumPy array of same shape;
e.g., `lambda x: x ** 2`, `lambda x: x / x.shape[0]` or `np.log`.
It can also be a function which takes a timestamp and array, and returns a new array of same shape;
e.g., `lambda ts, x: x / ts.days_in_month`.
The type of `ts` is either `pd.Timestamp` (if the series is indexed with a DatetimeIndex),
or an integer otherwise (if the series is indexed with an RangeIndex).
Returns
-------
TimeSeries
A new TimeSeries instance
"""
if not isinstance(fn, Callable):
raise_log(TypeError("fn should be callable"), logger)
if isinstance(fn, np.ufunc):
if fn.nin == 1 and fn.nout == 1:
num_args = 1
elif fn.nin == 2 and fn.nout == 1:
num_args = 2
else:
raise_log(
ValueError(
"fn must have either one or two arguments and return a single value"
),
logger,
)
else:
try:
num_args = len(signature(fn).parameters)
except ValueError:
raise_log(
ValueError(
"inspect.signature(fn) failed. Try wrapping fn in a lambda, e.g. lambda x: fn(x)"
),
logger,
)
new_xa = self._xa.copy()
if num_args == 1: # apply fn on values directly
new_xa.values = fn(self._xa.values)
elif num_args == 2: # map function uses timestamp f(timestamp, x)
# go over shortest amount of iterations, either over time steps or components and samples
if self.n_timesteps <= self.n_components * self.n_samples:
new_vals = np.vstack(
[
np.expand_dims(
fn(self.time_index[i], self._xa[i, :, :]), axis=0
)
for i in range(self.n_timesteps)
]
)
else:
new_vals = np.stack(
[
np.column_stack(
[
fn(self.time_index, self._xa[:, i, j])
for j in range(self.n_samples)
]
)
for i in range(self.n_components)
],
axis=1,
)
new_xa.values = new_vals
else:
raise_log(ValueError("fn must have either one or two arguments"), logger)
return self.__class__(new_xa)
[docs] def to_json(self) -> str:
"""
Return a JSON string representation of this deterministic series.
At the moment this function works only on deterministic time series (i.e., made of 1 sample).
Notes
-----
Static covariates are not returned in the JSON string. When using `TimeSeries.from_json()`, the static
covariates can be added with input argument `static_covariates`.
Returns
-------
str
A JSON String representing the time series
"""
return self.pd_dataframe().to_json(orient="split", date_format="iso")
[docs] def to_csv(self, *args, **kwargs):
"""
Writes this deterministic series to a CSV file.
For a list of parameters, refer to the documentation of :func:`pandas.DataFrame.to_csv()` [1]_.
References
----------
.. [1] https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.to_csv.html?highlight=to_csv
"""
if not self.is_deterministic:
raise_log(
AssertionError(
"The pd_dataframe() method can only return DataFrames of deterministic "
"time series, and this series is not deterministic (it contains several samples). "
"Consider calling quantile_df() instead."
)
)
self.pd_dataframe().to_csv(*args, **kwargs)
[docs] def to_pickle(self, path: str, protocol: int = pickle.HIGHEST_PROTOCOL):
"""
Save this series in pickle format.
Parameters
----------
path : string
path to a file where current object will be pickled
protocol : integer, default highest
pickling protocol. The default is best in most cases, use it only if having backward compatibility issues
Notes
-----
Xarray docs [1]_ suggest not using pickle as a long-term data storage.
References
----------
.. [1] http://xarray.pydata.org/en/stable/user-guide/io.html#pickle
"""
with open(path, "wb") as fh:
pickle.dump(self, fh, protocol=protocol)
[docs] def plot(
self,
new_plot: bool = False,
central_quantile: Union[float, str] = 0.5,
low_quantile: Optional[float] = 0.05,
high_quantile: Optional[float] = 0.95,
default_formatting: bool = True,
label: Optional[Union[str, Sequence[str]]] = "",
max_nr_components: int = 10,
ax: Optional[matplotlib.axes.Axes] = None,
*args,
**kwargs,
) -> matplotlib.axes.Axes:
"""Plot the series.
This is a wrapper method around :func:`xarray.DataArray.plot()`.
Parameters
----------
new_plot
Whether to spawn a new axis to plot on. See also parameter `ax`.
central_quantile
The quantile (between 0 and 1) to plot as a "central" value, if the series is stochastic (i.e., if
it has multiple samples). This will be applied on each component separately (i.e., to display quantiles
of the components' marginal distributions). For instance, setting `central_quantile=0.5` will plot the
median of each component. `central_quantile` can also be set to 'mean'.
low_quantile
The quantile to use for the lower bound of the plotted confidence interval. Similar to `central_quantile`,
this is applied to each component separately (i.e., displaying marginal distributions). No confidence
interval is shown if `confidence_low_quantile` is None (default 0.05).
high_quantile
The quantile to use for the upper bound of the plotted confidence interval. Similar to `central_quantile`,
this is applied to each component separately (i.e., displaying marginal distributions). No confidence
interval is shown if `high_quantile` is None (default 0.95).
default_formatting
Whether to use the darts default scheme.
label
Can either be a string or list of strings. If a string and the series only has a single component, it is
used as the label for that component. If a string and the series has multiple components, it is used as
a prefix for each component name. If a list of strings with length equal to the number of components in
the series, the labels will be mapped to the components in order.
max_nr_components
The maximum number of components of a series to plot. -1 means all components will be plotted.
ax
Optionally, an axis to plot on. If `None`, and `new_plot=False`, will use the current axis. If
`new_plot=True`, will create a new axis.
alpha
Optionally, set the line alpha for deterministic series, or the confidence interval alpha for
probabilistic series.
args
some positional arguments for the `plot()` method
kwargs
some keyword arguments for the `plot()` method
Returns
-------
matplotlib.axes.Axes
Either the passed `ax` axis, a newly created one if `new_plot=True`, or the existing one.
"""
alpha_confidence_intvls = 0.25
if central_quantile != "mean":
raise_if_not(
isinstance(central_quantile, float) and 0.0 <= central_quantile <= 1.0,
'central_quantile must be either "mean", or a float between 0 and 1.',
logger,
)
if high_quantile is not None and low_quantile is not None:
raise_if_not(
0.0 <= low_quantile <= 1.0 and 0.0 <= high_quantile <= 1.0,
"confidence interval low and high quantiles must be between 0 and 1.",
logger,
)
if new_plot:
fig, ax = plt.subplots()
else:
if ax is None:
ax = plt.gca()
if not any(lw in kwargs for lw in ["lw", "linewidth"]):
kwargs["lw"] = 2
n_components_to_plot = max_nr_components
if n_components_to_plot == -1:
n_components_to_plot = self.n_components
elif self.n_components > max_nr_components:
logger.warning(
f"Number of components is larger than {max_nr_components} ({self.n_components}). "
f"Plotting only the first {max_nr_components} components."
f"You can overwrite this in the using the `plot_all_components` argument in plot()"
f"Beware that plotting a large number of components may cause performance issues."
)
if not isinstance(label, str) and isinstance(label, Sequence):
raise_if_not(
len(label) == self.n_components
or (
self.n_components > n_components_to_plot
and len(label) >= n_components_to_plot
),
"The label argument should have the same length as the number of plotted components "
f"({min(self.n_components, n_components_to_plot)}), only {len(label)} labels were provided",
logger,
)
custom_labels = True
else:
custom_labels = False
for i, c in enumerate(self._xa.component[:n_components_to_plot]):
comp_name = str(c.values)
comp = self._xa.sel(component=c)
if comp.sample.size > 1:
if central_quantile == "mean":
central_series = comp.mean(dim=DIMS[2])
else:
central_series = comp.quantile(q=central_quantile, dim=DIMS[2])
else:
central_series = comp.mean(dim=DIMS[2])
alpha = kwargs["alpha"] if "alpha" in kwargs else None
if not self.is_deterministic:
kwargs["alpha"] = 1
if custom_labels:
label_to_use = label[i]
else:
if label == "":
label_to_use = comp_name
elif len(self.components) == 1:
label_to_use = label
else:
label_to_use = f"{label}_{comp_name}"
kwargs["label"] = label_to_use
if central_series.shape[0] > 1:
p = central_series.plot(*args, ax=ax, **kwargs)
# empty TimeSeries
elif central_series.shape[0] == 0:
p = ax.plot(
[],
[],
*args,
**kwargs,
)
ax.set_xlabel(self.time_index.name)
else:
p = ax.plot(
[self.start_time()],
central_series.values[0],
"o",
*args,
**kwargs,
)
color_used = p[0].get_color() if default_formatting else None
# Optionally show confidence intervals
if (
comp.sample.size > 1
and low_quantile is not None
and high_quantile is not None
):
low_series = comp.quantile(q=low_quantile, dim=DIMS[2])
high_series = comp.quantile(q=high_quantile, dim=DIMS[2])
if low_series.shape[0] > 1:
ax.fill_between(
self.time_index,
low_series,
high_series,
color=color_used,
alpha=(alpha if alpha is not None else alpha_confidence_intvls),
)
else:
ax.plot(
[self.start_time(), self.start_time()],
[low_series.values[0], high_series.values[0]],
"-+",
color=color_used,
lw=2,
)
ax.legend()
ax.set_title(self._xa.name)
return ax
[docs] def with_columns_renamed(
self, col_names: Union[List[str], str], col_names_new: Union[List[str], str]
) -> Self:
"""
Return a new ``TimeSeries`` instance with new columns/components names. It also
adapts the names in the hierarchy, if any.
Parameters
-------
col_names
String or list of strings corresponding the the column names to be changed.
col_names_new
String or list of strings corresponding to the new column names. Must be the same length as col_names.
Returns
-------
TimeSeries
A new TimeSeries instance.
"""
if isinstance(col_names, str):
col_names = [col_names]
if isinstance(col_names_new, str):
col_names_new = [col_names_new]
raise_if_not(
all([(x in self.columns.to_list()) for x in col_names]),
"Some column names in col_names don't exist in the time series.",
logger,
)
raise_if_not(
len(col_names) == len(col_names_new),
"Length of col_names_new list should be"
" equal to the length of col_names list.",
logger,
)
old2new = {old: new for (old, new) in zip(col_names, col_names_new)}
# update component names
cols = [old2new[old] if old in old2new else old for old in self.components]
# update hierarchy names
if self.hierarchy is not None:
hierarchy = {
(old2new[key] if key in old2new else key): [
old2new[old] if old in old2new else old
for old in self.hierarchy[key]
]
for key in self.hierarchy
}
else:
hierarchy = None
new_attrs = self._xa.attrs
new_attrs[HIERARCHY_TAG] = hierarchy
new_xa = xr.DataArray(
self._xa.values,
dims=self._xa.dims,
coords={self._xa.dims[0]: self.time_index, DIMS[1]: pd.Index(cols)},
attrs=new_attrs,
)
return self.__class__(new_xa)
"""
Simple statistic and aggregation functions. Calculate various statistics over the samples of stochastic time series
or aggregate over components/time for deterministic series.
"""
def _get_agg_coords(self, new_cname: str, axis: int) -> dict:
"""Helper function to rename reduced axis. Returns a dictionary containing the new coordinates"""
if axis == 0: # set time_index to first day
return {self._xa.dims[0]: self.time_index[0:1], DIMS[1]: self.components}
elif axis == 1: # rename components
return {self._xa.dims[0]: self.time_index, DIMS[1]: pd.Index([new_cname])}
elif axis == 2: # do nothing
return {self._xa.dims[0]: self.time_index, DIMS[1]: self.components}
[docs] def mean(self, axis: int = 2) -> Self:
"""
Return a ``TimeSeries`` containing the mean calculated over the specified axis.
If we reduce over time (``axis=0``), the resulting ``TimeSeries`` will have length one and will use the first
entry of the original ``time_index``. If we perform the calculation over the components (``axis=1``), the
resulting single component will be renamed to "components_mean". When applied to the samples (``axis=2``),
a deterministic ``TimeSeries`` is returned.
If ``axis=1``, the static covariates and the hierarchy are discarded from the series.
Parameters
----------
axis
The axis to reduce over. The default is to calculate over samples, i.e. axis=2.
Returns
-------
TimeSeries
A new TimeSeries with mean applied to the indicated axis.
"""
new_data = self._xa.values.mean(axis=axis, keepdims=True)
new_coords = self._get_agg_coords("components_mean", axis)
new_xa = xr.DataArray(
new_data,
dims=self._xa.dims,
coords=new_coords,
attrs=(self._xa.attrs if axis != 1 else dict()),
)
return self.__class__(new_xa)
[docs] def sum(self, axis: int = 2) -> Self:
"""
Return a ``TimeSeries`` containing the sum calculated over the specified axis.
If we reduce over time (``axis=0``), the resulting ``TimeSeries`` will have length one and will use the first
entry of the original ``time_index``. If we perform the calculation over the components (``axis=1``), the
resulting single component will be renamed to "components_sum". When applied to the samples (``axis=2``),
a deterministic ``TimeSeries`` is returned.
If ``axis=1``, the static covariates and the hierarchy are discarded from the series.
Parameters
----------
axis
The axis to reduce over. The default is to calculate over samples, i.e. axis=2.
Returns
-------
TimeSeries
A new TimeSeries with sum applied to the indicated axis.
"""
new_data = self._xa.values.sum(axis=axis, keepdims=True)
new_coords = self._get_agg_coords("components_sum", axis)
new_xa = xr.DataArray(
new_data,
dims=self._xa.dims,
coords=new_coords,
attrs=(self._xa.attrs if axis != 1 else dict()),
)
return self.__class__(new_xa)
[docs] def min(self, axis: int = 2) -> Self:
"""
Return a ``TimeSeries`` containing the min calculated over the specified axis.
If we reduce over time (``axis=0``), the resulting ``TimeSeries`` will have length one and will use the first
entry of the original ``time_index``. If we perform the calculation over the components (``axis=1``), the
resulting single component will be renamed to "components_min". When applied to the samples (``axis=2``),
a deterministic ``TimeSeries`` is returned.
If ``axis=1``, the static covariates and the hierarchy are discarded from the series.
Parameters
----------
axis
The axis to reduce over. The default is to calculate over samples, i.e. axis=2.
Returns
-------
TimeSeries
A new TimeSeries with min applied to the indicated axis.
"""
new_data = self._xa.values.min(axis=axis, keepdims=True)
new_coords = self._get_agg_coords("components_min", axis)
new_xa = xr.DataArray(
new_data,
dims=self._xa.dims,
coords=new_coords,
attrs=(self._xa.attrs if axis != 1 else dict()),
)
return self.__class__(new_xa)
[docs] def max(self, axis: int = 2) -> Self:
"""
Return a ``TimeSeries`` containing the max calculated over the specified axis.
If we reduce over time (``axis=0``), the resulting ``TimeSeries`` will have length one and will use the first
entry of the original ``time_index``. If we perform the calculation over the components (``axis=1``), the
resulting single component will be renamed to "components_max". When applied to the samples (``axis=2``),
a deterministic ``TimeSeries`` is returned.
If ``axis=1``, the static covariates and the hierarchy are discarded from the series.
Parameters
----------
axis
The axis to reduce over. The default is to calculate over samples, i.e. axis=2.
Returns
-------
TimeSeries
A new TimeSeries with max applied to the indicated axis.
"""
new_data = self._xa.values.max(axis=axis, keepdims=True)
new_coords = self._get_agg_coords("components_max", axis)
new_xa = xr.DataArray(
new_data,
dims=self._xa.dims,
coords=new_coords,
attrs=(self._xa.attrs if axis != 1 else dict()),
)
return self.__class__(new_xa)
[docs] def var(self, ddof: int = 1) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the variance of each component
(over the samples) of this stochastic ``TimeSeries``.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
ddof
"Delta Degrees of Freedom": the divisor used in the calculation is N - ddof where N represents the
number of elements. By default, ddof is 1.
Returns
-------
TimeSeries
The TimeSeries containing the variance for each component.
"""
self._assert_stochastic()
new_data = self._xa.values.var(axis=2, ddof=ddof, keepdims=True)
new_xa = xr.DataArray(
new_data, dims=self._xa.dims, coords=self._xa.coords, attrs=self._xa.attrs
)
return self.__class__(new_xa)
[docs] def std(self, ddof: int = 1) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the standard deviation of each component
(over the samples) of this stochastic ``TimeSeries``.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
ddof
"Delta Degrees of Freedom": the divisor used in the calculation is N - ddof where N represents the
number of elements. By default, ddof is 1.
Returns
-------
TimeSeries
The TimeSeries containing the standard deviation for each component.
"""
self._assert_stochastic()
new_data = self._xa.values.std(axis=2, ddof=ddof, keepdims=True)
new_xa = xr.DataArray(
new_data, dims=self._xa.dims, coords=self._xa.coords, attrs=self._xa.attrs
)
return self.__class__(new_xa)
[docs] def skew(self, **kwargs) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the skew of each component
(over the samples) of this stochastic ``TimeSeries``.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
kwargs
Other keyword arguments are passed down to `scipy.stats.skew()`
Returns
-------
TimeSeries
The TimeSeries containing the skew for each component.
"""
self._assert_stochastic()
new_data = np.expand_dims(skew(self._xa.values, axis=2, **kwargs), axis=2)
new_xa = xr.DataArray(
new_data, dims=self._xa.dims, coords=self._xa.coords, attrs=self._xa.attrs
)
return self.__class__(new_xa)
[docs] def kurtosis(self, **kwargs) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the kurtosis of each component
(over the samples) of this stochastic ``TimeSeries``.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
kwargs
Other keyword arguments are passed down to `scipy.stats.kurtosis()`
Returns
-------
TimeSeries
The TimeSeries containing the kurtosis for each component.
"""
self._assert_stochastic()
new_data = np.expand_dims(kurtosis(self._xa.values, axis=2, **kwargs), axis=2)
new_xa = xr.DataArray(
new_data, dims=self._xa.dims, coords=self._xa.coords, attrs=self._xa.attrs
)
return self.__class__(new_xa)
[docs] def quantile(self, quantile: float, **kwargs) -> Self:
"""
Return a deterministic ``TimeSeries`` containing the single desired quantile of each component
(over the samples) of this stochastic ``TimeSeries``.
The components in the new series are named "<component>_X", where "<component>"
is the column name corresponding to this component, and "X" is the quantile value.
The quantile columns represent the marginal distributions of the components of this series.
This works only on stochastic series (i.e., with more than 1 sample)
Parameters
----------
quantile
The desired quantile value. The value must be represented as a fraction
(between 0 and 1 inclusive). For instance, `0.5` will return a TimeSeries
containing the median of the (marginal) distribution of each component.
kwargs
Other keyword arguments are passed down to `numpy.quantile()`
Returns
-------
TimeSeries
The TimeSeries containing the desired quantile for each component.
"""
return self.quantile_timeseries(quantile, **kwargs)
"""
Dunder methods
"""
def _combine_arrays(
self,
other: Union[Self, xr.DataArray, np.ndarray],
combine_fn: Callable[[np.ndarray, np.ndarray], np.ndarray],
) -> Self:
"""
This is a helper function that allows us to combine this series with another one,
directly applying an operation on their underlying numpy arrays.
"""
if isinstance(other, TimeSeries):
other_vals = other.data_array(copy=False).values
elif isinstance(other, xr.DataArray):
other_vals = other.values
else:
other_vals = other
raise_if_not(
self._xa.values.shape == other_vals.shape,
"Attempted to perform operation on two TimeSeries of unequal shapes.",
logger,
)
new_xa = self._xa.copy()
new_xa.values = combine_fn(new_xa.values, other_vals)
return self.__class__(new_xa)
@classmethod
def _fill_missing_dates(
cls, xa: xr.DataArray, freq: Optional[Union[str, int]] = None
) -> xr.DataArray:
"""Return an xarray DataArray instance with missing dates inserted from an input xarray DataArray.
The first dimension of the input DataArray `xa` has to be the time dimension.
This requires either a provided `freq` or the possibility to infer a unique frequency (see
`offset aliases <https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_
for more info on supported frequencies) from the provided timestamps.
Parameters
----------
xa
The xarray DataArray
freq
Optionally, a string representing the frequency of the Pandas DateTimeIndex to fill in the missing dates.
Raises
-------
ValueError
If `xa`'s DateTimeIndex contains less than 3 elements;
if no unique frequency can be inferred from `xa`'s DateTimeIndex;
if the resampled DateTimeIndex does not contain all dates from `xa` (see
:meth:`_restore_xarray_from_frequency() <TimeSeries._restore_xarray_from_frequency>`)
Returns
-------
xarray DataArray
xarray DataArray with filled missing dates from `xa`
"""
if freq is not None:
return cls._restore_xarray_from_frequency(xa, freq)
raise_if(
len(xa) <= 2,
"Input time series must be of (length>=3) when fill_missing_dates=True and freq=None.",
logger,
)
time_dim = xa.dims[0]
sorted_xa = cls._sort_index(xa, copy=False)
time_index: Union[pd.Index, pd.RangeIndex, pd.DatetimeIndex] = (
sorted_xa.get_index(time_dim)
)
if isinstance(time_index, pd.DatetimeIndex):
has_datetime_index = True
observed_freqs = cls._observed_freq_datetime_index(time_index)
else: # integer index (non RangeIndex)
has_datetime_index = False
observed_freqs = cls._observed_freq_integer_index(time_index)
offset_alias_info = (
(
" For more information about frequency aliases, read "
"https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases"
)
if has_datetime_index
else ""
)
if not len(observed_freqs) == 1:
raise_log(
ValueError(
f"Could not observe an inferred frequency. An explicit frequency must be evident over a span of "
f"at least 3 consecutive time stamps in the input data. {offset_alias_info}"
if not len(observed_freqs)
else f"Could not find a unique inferred frequency (not constant). Observed frequencies: "
f"{observed_freqs}. If any of those is the actual frequency, try passing it with "
f"`fill_missing_dates=True` and `freq=your_frequency`.{offset_alias_info}"
),
logger,
)
freq = observed_freqs.pop()
return cls._restore_xarray_from_frequency(sorted_xa, freq)
@staticmethod
def _sort_index(xa: xr.DataArray, copy: bool = True) -> xr.DataArray:
"""Sorts an xarray by its time dimension index (only if it is not already monotonically increasing)."""
time_dim = xa.dims[0]
return (
(xa.copy() if copy else xa)
if xa.get_index(time_dim).is_monotonic_increasing
else xa.sortby(time_dim)
)
@staticmethod
def _observed_freq_datetime_index(index: pd.DatetimeIndex) -> set:
"""Returns all observed/inferred frequencies of a pandas DatetimeIndex. The frequencies are inferred from all
combinations of three adjacent time steps
"""
# find unique time deltas indices from three consecutive time stamps
_, unq_td_index = np.unique(
np.stack([(index[1:-1] - index[:-2]), (index[2:] - index[1:-1])]),
return_index=True,
axis=1,
)
# for each unique index, take one example including the left time stamp, and one including the right
steps = np.column_stack([index[unq_td_index + i] for i in range(3)])
# find all unique inferred frequencies
observed_freqs = {pd.infer_freq(step) for step in steps}
observed_freqs.discard(None)
return observed_freqs
@staticmethod
def _observed_freq_integer_index(index: pd.Index) -> set:
"""Returns all observed/inferred frequencies of a pandas Index (integer index). The frequencies are inferred
from all differences between two adjacent indices.
"""
return set(index[1:] - index[:-1])
@classmethod
def _integer_to_range_indexed_xarray(cls, xa: xr.DataArray) -> xr.DataArray:
"""If possible, converts an integer indexed xarray DataArray to a range indexed (pd.RangeIndex) DataArray.
Otherwise, raises an error. An integer Index can be converted to a pd.RangeIndex, if the sorted integer index
has a constant step size.
"""
time_dim = xa.dims[0]
sorted_xa = cls._sort_index(xa, copy=False)
time_index = sorted_xa.get_index(time_dim)
observed_freqs = cls._observed_freq_integer_index(time_index)
raise_if_not(
len(observed_freqs) == 1,
f"Could not convert integer index to a pd.RangeIndex. Found non-unique step sizes/frequencies: "
f"{observed_freqs}. If any of those is the actual frequency, try passing it with fill_missing_dates=True "
f"and freq=your_frequency.",
logger,
)
freq = observed_freqs.pop()
idx = pd.RangeIndex(
start=min(time_index),
stop=max(time_index) + freq,
step=freq,
name=time_index.name,
)
coords = {
str(xa.dims[0]): idx,
str(xa.dims[1]): xa.coords[DIMS[1]],
}
return xr.DataArray(
data=sorted_xa.data,
dims=xa.dims,
coords=coords,
attrs=xa.attrs,
)
@classmethod
def _restore_xarray_from_frequency(
cls, xa: xr.DataArray, freq: Union[str, int]
) -> xr.DataArray:
"""Return an xarray DataArray instance that is resampled from an input xarray DataArray `xa` with frequency
`freq`. `freq` should be the inferred or actual frequency of `xa`. All data from `xa` is maintained in the
output DataArray at the corresponding dates. Any missing dates from `xa` will be inserted into the returned
DataArray with np.nan values.
The first dimension of the input DataArray `xa` has to be the time dimension.
This requires a provided frequency/step size `freq`
Parameters
----------
xa
The xarray DataArray
freq
If a string, represents the actual or inferred frequency of the pandas DatetimeIndex from `xa` (see
`offset aliases <https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases>`_
for more info on supported frequencies).
If an integer, represents the actual or inferred step size of the pandas Index or pandas RangeIndex from
`xa`.
Raises
-------
ValueError
If the resampled/reindexed DateTimeIndex/RangeIndex does not contain all dates from `xa`
Returns
-------
xarray DataArray
xarray DataArray resampled from `xa` with `freq` including all data from `xa` and inserted missing dates
"""
time_dim = xa.dims[0]
sorted_xa = cls._sort_index(xa, copy=False)
time_index = sorted_xa.get_index(time_dim)
resampled_time_index = pd.Series(index=time_index, dtype="object")
if isinstance(time_index, pd.DatetimeIndex):
has_datetime_index = True
resampled_time_index = resampled_time_index.asfreq(freq)
else: # integer index (non RangeIndex) -> resampled to RangeIndex
has_datetime_index = False
resampled_time_index = resampled_time_index.reindex(
range(min(time_index), max(time_index) + freq, freq)
)
# check if new time index with inferred frequency contains all input data
contains_all_data = time_index.isin(resampled_time_index.index).all()
offset_alias_info = (
(
" For more information about frequency aliases, read "
"https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases"
)
if has_datetime_index
else ""
)
if not contains_all_data:
raise_log(
ValueError(
f"Could not correctly fill missing {'dates' if has_datetime_index else 'indices'} with the "
f"observed/passed {'frequency' if has_datetime_index else 'step size'} `freq='{freq}'`. "
f"Not all input {'time stamps' if has_datetime_index else 'indices'} contained in the newly "
f"created TimeSeries.{offset_alias_info}"
),
logger,
)
coords = {
str(xa.dims[0]): resampled_time_index.index,
str(xa.dims[1]): xa.coords[DIMS[1]],
}
# convert to float as for instance integer arrays cannot accept nans
dtype = (
xa.dtype
if (
np.issubdtype(xa.values.dtype, np.float32)
or np.issubdtype(xa.values.dtype, np.float64)
)
else np.float64
)
resampled_xa = xr.DataArray(
data=np.empty(
shape=((len(resampled_time_index),) + xa.shape[1:]), dtype=dtype
),
dims=xa.dims,
coords=coords,
attrs=xa.attrs,
)
resampled_xa[:] = np.nan
resampled_xa[resampled_time_index.index.isin(time_index)] = sorted_xa.data
return resampled_xa
def _get_dim_name(self, axis: Union[int, str]) -> str:
if isinstance(axis, int):
if axis == 0:
return self._time_dim
elif axis == 1 or axis == 2:
return DIMS[axis]
else:
raise_if(True, "If `axis` is an integer it must be between 0 and 2.")
else:
known_dims = (self._time_dim,) + DIMS[1:]
raise_if_not(
axis in known_dims,
"`axis` must be a known dimension of this series: {}".format(
known_dims
),
)
return axis
def _get_dim(self, axis: Union[int, str]) -> int:
if isinstance(axis, int):
raise_if_not(
0 <= axis <= 2, "If `axis` is an integer it must be between 0 and 2."
)
return axis
else:
known_dims = (self._time_dim,) + DIMS[1:]
raise_if_not(
axis in known_dims,
"`axis` must be a known dimension of this series: {}".format(
known_dims
),
)
return known_dims.index(axis)
def __eq__(self, other):
if isinstance(other, TimeSeries):
return self._xa.equals(other.data_array(copy=False))
return False
def __ne__(self, other):
return not self.__eq__(other)
def __len__(self):
return len(self._xa)
def __add__(self, other):
if isinstance(other, (int, float, np.integer)):
xa_ = _xarray_with_attrs(
self._xa + other, self.static_covariates, self.hierarchy
)
return self.__class__(xa_)
elif isinstance(other, (TimeSeries, xr.DataArray, np.ndarray)):
return self._combine_arrays(other, lambda s1, s2: s1 + s2)
else:
raise_log(
TypeError(
"unsupported operand type(s) for + or add(): '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __radd__(self, other):
return self + other
def __sub__(self, other):
if isinstance(other, (int, float, np.integer)):
xa_ = _xarray_with_attrs(
self._xa - other, self.static_covariates, self.hierarchy
)
return self.__class__(xa_)
elif isinstance(other, (TimeSeries, xr.DataArray, np.ndarray)):
return self._combine_arrays(other, lambda s1, s2: s1 - s2)
else:
raise_log(
TypeError(
"unsupported operand type(s) for - or sub(): '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __rsub__(self, other):
return other + (-self)
def __mul__(self, other):
if isinstance(other, (int, float, np.integer)):
xa_ = _xarray_with_attrs(
self._xa * other, self.static_covariates, self.hierarchy
)
return self.__class__(xa_)
elif isinstance(other, (TimeSeries, xr.DataArray, np.ndarray)):
return self._combine_arrays(other, lambda s1, s2: s1 * s2)
else:
raise_log(
TypeError(
"unsupported operand type(s) for * or mul(): '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __rmul__(self, other):
return self * other
def __pow__(self, n):
if isinstance(n, (int, float, np.integer)):
raise_if(n < 0, "Attempted to raise a series to a negative power.", logger)
xa_ = _xarray_with_attrs(
self._xa ** float(n), self.static_covariates, self.hierarchy
)
return self.__class__(xa_)
if isinstance(n, (TimeSeries, xr.DataArray, np.ndarray)):
return self._combine_arrays(n, lambda s1, s2: s1**s2) # elementwise power
else:
raise_log(
TypeError(
"unsupported operand type(s) for ** or pow(): '{}' and '{}'.".format(
type(self).__name__, type(n).__name__
)
),
logger,
)
def __truediv__(self, other):
if isinstance(other, (int, float, np.integer)):
if other == 0:
raise_log(ZeroDivisionError("Cannot divide by 0."), logger)
xa_ = _xarray_with_attrs(
self._xa / other, self.static_covariates, self.hierarchy
)
return self.__class__(xa_)
elif isinstance(other, (TimeSeries, xr.DataArray, np.ndarray)):
if not (other.all_values(copy=False) != 0).all():
raise_log(
ZeroDivisionError("Cannot divide by a TimeSeries with a value 0."),
logger,
)
return self._combine_arrays(other, lambda s1, s2: s1 / s2)
else:
raise_log(
TypeError(
"unsupported operand type(s) for / or truediv(): '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __rtruediv__(self, n):
return n * (self ** (-1))
def __abs__(self):
return self.__class__(abs(self._xa))
def __neg__(self):
return self.__class__(-self._xa)
def __contains__(self, ts: Union[int, pd.Timestamp]) -> bool:
return ts in self.time_index
def __round__(self, n=None):
return self.__class__(self._xa.round(n))
def __lt__(self, other) -> xr.DataArray:
if isinstance(other, (int, float, np.integer, np.ndarray, xr.DataArray)):
return _xarray_with_attrs(
self._xa < other, self.static_covariates, self.hierarchy
)
elif isinstance(other, TimeSeries):
return _xarray_with_attrs(
self._xa < other.data_array(copy=False),
self.static_covariates,
self.hierarchy,
)
else:
raise_log(
TypeError(
"unsupported operand type(s) for < : '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __gt__(self, other) -> xr.DataArray:
if isinstance(other, (int, float, np.integer, np.ndarray, xr.DataArray)):
return _xarray_with_attrs(
self._xa > other, self.static_covariates, self.hierarchy
)
elif isinstance(other, TimeSeries):
return _xarray_with_attrs(
self._xa > other.data_array(copy=False),
self.static_covariates,
self.hierarchy,
)
else:
raise_log(
TypeError(
"unsupported operand type(s) for < : '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __le__(self, other) -> xr.DataArray:
if isinstance(other, (int, float, np.integer, np.ndarray, xr.DataArray)):
return _xarray_with_attrs(
self._xa <= other, self.static_covariates, self.hierarchy
)
elif isinstance(other, TimeSeries):
return _xarray_with_attrs(
self._xa <= other.data_array(copy=False),
self.static_covariates,
self.hierarchy,
)
else:
raise_log(
TypeError(
"unsupported operand type(s) for < : '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __ge__(self, other) -> xr.DataArray:
if isinstance(other, (int, float, np.integer, np.ndarray, xr.DataArray)):
return _xarray_with_attrs(
self._xa >= other, self.static_covariates, self.hierarchy
)
elif isinstance(other, TimeSeries):
return _xarray_with_attrs(
self._xa >= other.data_array(copy=False),
self.static_covariates,
self.hierarchy,
)
else:
raise_log(
TypeError(
"unsupported operand type(s) for < : '{}' and '{}'.".format(
type(self).__name__, type(other).__name__
)
),
logger,
)
def __str__(self):
return str(self._xa).replace("xarray.DataArray", "TimeSeries (DataArray)")
def __repr__(self):
return self._xa.__repr__().replace("xarray.DataArray", "TimeSeries (DataArray)")
def _repr_html_(self):
return self._xa._repr_html_().replace(
"xarray.DataArray", "TimeSeries (DataArray)"
)
def __copy__(self, deep: bool = True):
return self.copy()
def __deepcopy__(self, memo):
return self.__class__(deepcopy(self._xa, memo))
def __getitem__(
self,
key: Union[
pd.DatetimeIndex,
pd.RangeIndex,
List[str],
List[int],
List[pd.Timestamp],
str,
int,
pd.Timestamp,
Any,
],
) -> Self:
"""Allow indexing on darts TimeSeries.
The supported index types are the following base types as a single value, a list or a slice:
- pd.Timestamp -> return a TimeSeries corresponding to the value(s) at the given timestamp(s).
- str -> return a TimeSeries including the column(s) (components) specified as str.
- int -> return a TimeSeries with the value(s) at the given row (time) index.
`pd.DatetimeIndex` and `pd.RangeIndex` are also supported and will return the corresponding value(s)
at the provided time indices.
.. warning::
slices use pandas convention of including both ends of the slice.
Notes
-----
For integer-indexed series, integers or slices of integer will return the result
of ``isel()``. That is, if integer ``i`` is provided, it returns the ``i``-th value
along the series, which is not necessarily the value where the time index is equal to ``i``
(e.g., if the time index does not start at 0). In contrast, calling this method with a
``pd.RangeIndex`` returns the result of ``sel()`` - i.e., the values where the time
index matches the provided range index.
"""
def _check_dt():
if not self._has_datetime_index:
raise_log(
ValueError(
"Attempted indexing a series with a DatetimeIndex or a timestamp, "
"but the series uses a RangeIndex."
),
logger,
)
def _check_range():
if self._has_datetime_index:
raise_log(
ValueError(
"Attempted indexing a series with a RangeIndex, "
"but the series uses a DatetimeIndex."
),
logger,
)
def _set_freq_in_xa(xa_in: xr.DataArray, freq=None):
# mutates the DataArray to make sure it contains the freq
if isinstance(xa_in.get_index(self._time_dim), pd.DatetimeIndex):
if freq is None:
freq = xa_in.get_index(self._time_dim).inferred_freq
if freq is not None:
xa_in.get_index(self._time_dim).freq = freq
else:
xa_in.get_index(self._time_dim).freq = self._freq
def _get_freq(xa_in: xr.DataArray):
if self._has_datetime_index:
return xa_in.get_index(self._time_dim).freq
else:
return xa_in.get_index(self._time_dim).step
adapt_covs_on_component = (
True
if self.has_static_covariates and len(self.static_covariates) > 1
else False
)
# handle DatetimeIndex and RangeIndex:
if isinstance(key, pd.DatetimeIndex):
_check_dt()
xa_ = self._xa.sel({self._time_dim: key})
# indexing may discard the freq, so we restore it...
# if the DateTimeIndex already has an associated freq, use it
# otherwise key.freq is None and the freq will be inferred
_set_freq_in_xa(xa_, key.freq)
return self.__class__(xa_)
elif isinstance(key, pd.RangeIndex):
_check_range()
idx_ = key
if not len(key) and self.freq != key.step:
# keep original step size in case of empty range index
idx_ = pd.RangeIndex(step=self.freq)
xa_ = self._xa.sel({self._time_dim: idx_})
# sel() gives us an Int64Index. We have to set the RangeIndex.
# see: https://github.com/pydata/xarray/issues/6256
xa_ = xa_.assign_coords({self.time_dim: idx_})
return self.__class__(xa_)
# handle slices:
elif isinstance(key, slice):
if isinstance(key.start, str) or isinstance(key.stop, str):
xa_ = self._xa.sel({DIMS[1]: key})
# selecting components discards the hierarchy, if any
xa_ = _xarray_with_attrs(
xa_,
(
xa_.attrs[STATIC_COV_TAG][key.start : key.stop]
if adapt_covs_on_component
else xa_.attrs[STATIC_COV_TAG]
),
None,
)
return self.__class__(xa_)
elif isinstance(key.start, (int, np.int64)) or isinstance(
key.stop, (int, np.int64)
):
xa_ = self._xa.isel({self._time_dim: key})
if _get_freq(xa_) is None:
# indexing discarded the freq; we restore it
freq = key.step * self.freq if key.step else self.freq
_set_freq_in_xa(xa_, freq)
return self.__class__(xa_)
elif isinstance(key.start, pd.Timestamp) or isinstance(
key.stop, pd.Timestamp
):
_check_dt()
xa_ = self._xa.sel({self._time_dim: key})
if _get_freq(xa_) is None:
# indexing discarded the freq; we restore it
freq = key.step * self.freq if key.step else self.freq
_set_freq_in_xa(xa_, freq)
return self.__class__(xa_)
# handle simple types:
elif isinstance(key, str):
# have to put key in a list not to drop the dimension
xa_ = self._xa.sel({DIMS[1]: [key]})
# selecting components discards the hierarchy, if any
xa_ = _xarray_with_attrs(
xa_,
(
xa_.attrs[STATIC_COV_TAG].loc[[key]]
if adapt_covs_on_component
else xa_.attrs[STATIC_COV_TAG]
),
None,
)
return self.__class__(xa_)
elif isinstance(key, (int, np.int64)):
xa_ = self._xa.isel({self._time_dim: [key]})
# restore a RangeIndex if needed:
time_idx = xa_.get_index(self._time_dim)
if pd.api.types.is_integer_dtype(time_idx) and not isinstance(
time_idx, pd.RangeIndex
):
xa_ = xa_.assign_coords(
{
self._time_dim: pd.RangeIndex(
start=time_idx[0],
stop=time_idx[0] + self.freq,
step=self.freq,
)
}
)
# indexing may discard the freq, so we restore it...
_set_freq_in_xa(xa_, freq=self.freq)
return self.__class__(xa_)
elif isinstance(key, pd.Timestamp):
_check_dt()
# indexing may discard the freq, so we restore it...
xa_ = self._xa.sel({self._time_dim: [key]})
_set_freq_in_xa(xa_, self.freq)
return self.__class__(xa_)
# handle lists:
if isinstance(key, list):
if all(isinstance(s, str) for s in key):
# when string(s) are provided, we consider it as (a list of) component(s)
xa_ = self._xa.sel({DIMS[1]: key})
xa_ = _xarray_with_attrs(
xa_,
(
xa_.attrs[STATIC_COV_TAG].loc[key]
if adapt_covs_on_component
else xa_.attrs[STATIC_COV_TAG]
),
None,
)
return self.__class__(xa_)
elif all(isinstance(i, (int, np.int64)) for i in key):
xa_ = self._xa.isel({self._time_dim: key})
# indexing may discard the freq, so we restore it...
_set_freq_in_xa(xa_)
orig_idx = self.time_index
if isinstance(orig_idx, pd.RangeIndex):
# We have to restore a RangeIndex. But first we need to
# check the list is corresponding to a RangeIndex.
min_idx, max_idx = min(key), max(key)
if (
not key[0] == min_idx
and key[-1] == max_idx
and max_idx + 1 - min_idx == len(key)
):
raise_log(
ValueError(
"Indexing a TimeSeries with a list requires the list to "
"contain monotonically increasing integers with no gap."
),
logger=logger,
)
new_idx = orig_idx[min_idx : max_idx + 1]
xa_ = xa_.assign_coords({self._time_dim: new_idx})
return self.__class__(xa_)
elif all(isinstance(t, pd.Timestamp) for t in key):
_check_dt()
# indexing may discard the freq, so we restore it...
xa_ = self._xa.sel({self._time_dim: key})
_set_freq_in_xa(xa_)
return self.__class__(xa_)
raise_log(IndexError("The type of your index was not matched."), logger)
def _xarray_with_attrs(xa_, static_covariates, hierarchy):
"""Return an DataArray instance with static covariates and hierarchy stored in the array's attributes.
Warning: This is an inplace operation (mutable) and should only be called from within TimeSeries construction
or to restore static covariates and hierarchy after operations in which they did not get transferred.
"""
xa_.attrs[STATIC_COV_TAG] = static_covariates
xa_.attrs[HIERARCHY_TAG] = hierarchy
return xa_
def _concat_static_covs(series: Sequence[TimeSeries]) -> Optional[pd.DataFrame]:
"""Concatenates static covariates along component dimension (rows of static covariates). For stacking or
concatenating TimeSeries along component dimension (axis=1).
Some context for stacking or concatenating two or more TimeSeries with static covariates:
Concat along axis=0 (time)
Along time dimension, we only take the static covariates of the first series (as static covariates are
time-independent).
Concat along axis=1 (components) or stacking
Along component dimension, we either concatenate or transfer the static covariates of the series if one
of below cases applies:
1) concatenate along component dimension (rows of static covariates) when for each series the number of
static covariate components is equal to the number of components in the series. The static variable
names (columns in series.static_covariates) must be identical across all series
2) if only the first series contains static covariates transfer only those
3) if `ignore_static_covarites=True` (with `concatenate()`), case 1) is ignored and only the static
covariates of the first series are transferred
Concat along axis=2 (samples)
Along sample dimension, we only take the static covariates of the first series (as we components and
time don't change).
"""
if not any([ts.has_static_covariates for ts in series]):
return None
only_first = series[0].has_static_covariates and not any(
[ts.has_static_covariates for ts in series[1:]]
)
all_have = all([ts.has_static_covariates for ts in series])
raise_if_not(
only_first or all_have,
"Either none, only the first or all TimeSeries must have `static_covariates`.",
logger,
)
if only_first:
return series[0].static_covariates
raise_if_not(
all([len(ts.static_covariates) == ts.n_components for ts in series])
and all(
[
ts.static_covariates.columns.equals(series[0].static_covariates.columns)
for ts in series
]
),
"Concatenation of multiple TimeSeries with static covariates requires all `static_covariates` "
"DataFrames to have identical columns (static variable names), and the number of each TimeSeries' "
"components must match the number of corresponding static covariate components (the number of rows "
"in `series.static_covariates`).",
logger,
)
return pd.concat(
[ts.static_covariates for ts in series if ts.has_static_covariates], axis=0
)
def _concat_hierarchy(series: Sequence[TimeSeries]):
"""
Used to concatenate the hierarchies of multiple TimeSeries, when concatenating series
along axis 1 (components). This simply merges the hierarchy dictionaries.
"""
concat_hierarchy = dict()
for s in series:
if s.has_hierarchy:
concat_hierarchy.update(s.hierarchy)
return None if len(concat_hierarchy) == 0 else concat_hierarchy
[docs]def concatenate(
series: Sequence[TimeSeries],
axis: Union[str, int] = 0,
ignore_time_axis: bool = False,
ignore_static_covariates: bool = False,
drop_hierarchy: bool = True,
):
"""Concatenates multiple ``TimeSeries`` along a given axis.
``axis`` can be an integer in (0, 1, 2) to denote (time, component, sample) or, alternatively,
a string denoting the corresponding dimension of the underlying ``DataArray``.
Parameters
----------
series : Sequence[TimeSeries]
sequence of ``TimeSeries`` to concatenate
axis : Union[str, int]
axis along which the series will be concatenated.
ignore_time_axis : bool
Allow concatenation even when some series do not have matching time axes.
When done along component or sample dimensions, concatenation will work as long as the series
have the same lengths (in this case the resulting series will have the time axis of the first
provided series). When done along time dimension, concatenation will work even if the time axes
are not contiguous (in this case, the resulting series will have a start time matching the start time
of the first provided series). Default: False.
ignore_static_covariates : bool
whether to ignore all requirements for static covariate concatenation and only transfer the static covariates
of the first TimeSeries element in `series` to the concatenated TimeSeries. Only effective when `axis=1`.
drop_hierarchy : bool
When `axis=1`, whether to drop hierarchy information. True by default. When False, the hierarchies will be
"concatenated" as well (by merging the hierarchy dictionaries), which may cause issues if the component
names of the resulting series and that of the merged hierarchy do not match.
When `axis=0` or `axis=2`, the hierarchy of the first series is always kept.
Returns
-------
TimeSeries
concatenated series
"""
time_dims = [ts.time_dim for ts in series]
if isinstance(axis, str):
if axis == DIMS[1]:
axis = 1
elif axis == DIMS[2]:
axis = 2
else:
raise_if_not(
len(set(time_dims)) == 1 and axis == time_dims[0],
"Unrecognised `axis` name. If `axis` denotes the time axis, all provided "
"series must have the same time axis name (if that is not the case, try providing "
"`axis=0` to concatenate along time dimension).",
)
axis = 0
# At this point all series are supposed to have same time dim name
time_dim_name = time_dims[0]
da_sequence = [ts.data_array(copy=False) for ts in series]
component_axis_equal = len({ts.width for ts in series}) == 1
sample_axis_equal = len({ts.n_samples for ts in series}) == 1
if axis == 0:
# time
raise_if(
(not (component_axis_equal and sample_axis_equal)),
"when concatenating along time dimension, the component and sample dimensions of all "
"provided series must match.",
)
da_concat = xr.concat(da_sequence, dim=time_dim_name)
# check, if timeseries are consecutive
consecutive_time_axes = True
for i in range(1, len(series)):
if series[i - 1].end_time() + series[0].freq != series[i].start_time():
consecutive_time_axes = False
break
if not consecutive_time_axes:
raise_if_not(
ignore_time_axis,
"When concatenating over time axis, all series need to be contiguous "
"in the time dimension. Use `ignore_time_axis=True` to override "
"this behavior and concatenate the series by extending the time axis "
"of the first series.",
)
tindex = generate_index(
start=series[0].start_time(),
freq=series[0].freq_str,
length=da_concat.shape[0],
)
da_concat = da_concat.assign_coords({time_dim_name: tindex})
da_concat = _xarray_with_attrs(
da_concat, series[0].static_covariates, series[0].hierarchy
)
else:
time_axes_equal = all(
list(
map(
lambda t: t[0].has_same_time_as(t[1]), zip(series[0:-1], series[1:])
)
)
)
time_axes_ok = (
time_axes_equal
if not ignore_time_axis
else len({len(ts) for ts in series}) == 1
)
raise_if_not(
(
time_axes_ok
and (
(axis == 1 and sample_axis_equal)
or (axis == 2 and component_axis_equal)
)
),
"When concatenating along component or sample dimensions, all the series must have the same time "
"axes (unless `ignore_time_axis` is True), or time axes of same lengths (if `ignore_time_axis` is "
"True), and all series must have the same number of samples (if concatenating along component "
"dimension), or the same number of components (if concatenating along sample dimension).",
)
# we concatenate raw values using Numpy because not all series might have the same time axes
# and joining using xarray.concatenate() won't work in some cases
concat_vals = np.concatenate([da.values for da in da_sequence], axis=axis)
if axis == 1:
# When concatenating along component dimension, we have to re-create a component index
# we rely on the factory method of TimeSeries to disambiguate names later on if needed.
component_index = pd.Index(
[c for cl in [ts.components for ts in series] for c in cl]
)
static_covariates = (
_concat_static_covs(series)
if not ignore_static_covariates
else series[0].static_covariates
)
hierarchy = None if drop_hierarchy else _concat_hierarchy(series)
else:
component_index = da_sequence[0].get_index(DIMS[1])
static_covariates = series[0].static_covariates
hierarchy = series[0].hierarchy
da_concat = xr.DataArray(
concat_vals,
dims=(time_dim_name,) + DIMS[-2:],
coords={time_dim_name: series[0].time_index, DIMS[1]: component_index},
attrs={STATIC_COV_TAG: static_covariates, HIERARCHY_TAG: hierarchy},
)
return TimeSeries.from_xarray(da_concat, fill_missing_dates=False)
def _finite_rows_boundaries(
values: np.ndarray, how: str = "all"
) -> Tuple[Optional[int], Optional[int]]:
"""
Return the indices of the first rows containing finite values starting from the start and the end of the first
dimension of the ndarray.
Parameters
----------
values
1D, 2D or 3D numpy array where the first dimension correspond to entries/rows, and the second to components/
columns
how
Define if the entries containing `NaN` in all the components ('all') or in any of the components ('any')
should be stripped. Default: 'all'
"""
dims = values.shape
raise_if(
len(dims) > 3, f"Expected 1D to 3D array, received {len(dims)}D array", logger
)
finite_rows = ~np.isnan(values)
if len(dims) == 3:
finite_rows = finite_rows.all(axis=2)
if len(dims) > 1 and dims[1] > 1:
if how == "any":
finite_rows = finite_rows.all(axis=1)
elif how == "all":
finite_rows = finite_rows.any(axis=1)
else:
raise_log(
ValueError(
f"`how` parameter value not recognized, should be either 'all' or 'any', "
f"received {how}"
)
)
first_finite_row = finite_rows.argmax()
last_finite_row = len(finite_rows) - finite_rows[::-1].argmax() - 1
return first_finite_row, last_finite_row