# -*- coding: utf-8 -*-
import copy
import time
import warnings
import pandas as pd
import numpy as np
from sklearn.metrics import mean_squared_error, mean_absolute_error
from sklearn.metrics.pairwise import euclidean_distances
from sklearn import preprocessing
from tsam.periodAggregation import aggregatePeriods
from tsam.representations import representations
pd.set_option("mode.chained_assignment", None)
# max iterator while resacling cluster profiles
MAX_ITERATOR = 20
# tolerance while rescaling cluster periods to meet the annual sum of the original profile
TOLERANCE = 1e-6
# minimal weight that overwrites a weighting of zero in order to carry the profile through the aggregation process
MIN_WEIGHT = 1e-6
[docs]def unstackToPeriods(timeSeries, timeStepsPerPeriod):
"""
Extend the timeseries to an integer multiple of the period length and
groups the time series to the periods.
:param timeSeries:
:type timeSeries: pandas DataFrame
:param timeStepsPerPeriod: The number of discrete timesteps which describe one period. required
:type timeStepsPerPeriod: integer
:returns: - **unstackedTimeSeries** (pandas DataFrame) -- is stacked such that each row represents a
candidate period
- **timeIndex** (pandas Series index) -- is the modification of the original
timeseriesindex in case an integer multiple was created
"""
# init new grouped timeindex
unstackedTimeSeries = timeSeries.copy()
# initialize new indices
periodIndex = []
stepIndex = []
# extend to inger multiple of period length
if len(timeSeries) % timeStepsPerPeriod == 0:
attached_timesteps = 0
else:
# calculate number of timesteps which get attached
attached_timesteps = timeStepsPerPeriod - len(timeSeries) % timeStepsPerPeriod
# take these from the head of the original time series
rep_data = unstackedTimeSeries.head(attached_timesteps)
# append them at the end of the time series
unstackedTimeSeries = pd.concat([unstackedTimeSeries, rep_data])
# create period and step index
for ii in range(0, len(unstackedTimeSeries)):
periodIndex.append(int(ii / timeStepsPerPeriod))
stepIndex.append(ii - int(ii / timeStepsPerPeriod) * timeStepsPerPeriod)
# save old index
timeIndex = copy.deepcopy(unstackedTimeSeries.index)
# create new double index and unstack the time series
unstackedTimeSeries.index = pd.MultiIndex.from_arrays(
[stepIndex, periodIndex], names=["TimeStep", "PeriodNum"]
)
unstackedTimeSeries = unstackedTimeSeries.unstack(level="TimeStep")
return unstackedTimeSeries, timeIndex
[docs]class TimeSeriesAggregation(object):
"""
Clusters time series data to typical periods.
"""
CLUSTER_METHODS = [
"averaging",
"k_means",
"k_medoids",
"k_maxoids",
"hierarchical",
"adjacent_periods",
]
REPRESENTATION_METHODS = [
"meanRepresentation",
"medoidRepresentation",
"maxoidRepresentation",
"minmaxmeanRepresentation",
"durationRepresentation",
"distributionRepresentation",
"distributionAndMinMaxRepresentation",
]
EXTREME_PERIOD_METHODS = [
"None",
"append",
"new_cluster_center",
"replace_cluster_center",
]
[docs] def __init__(
self,
timeSeries,
resolution=None,
noTypicalPeriods=10,
noSegments=10,
hoursPerPeriod=24,
clusterMethod="hierarchical",
evalSumPeriods=False,
sortValues=False,
sameMean=False,
rescaleClusterPeriods=True,
weightDict=None,
segmentation=False,
extremePeriodMethod="None",
representationMethod=None,
representationDict=None,
distributionPeriodWise=True,
segmentRepresentationMethod=None,
predefClusterOrder=None,
predefClusterCenterIndices=None,
solver="highs",
roundOutput=None,
addPeakMin=None,
addPeakMax=None,
addMeanMin=None,
addMeanMax=None,
):
"""
Initialize the periodly clusters.
:param timeSeries: DataFrame with the datetime as index and the relevant
time series parameters as columns. required
:type timeSeries: pandas.DataFrame() or dict
:param resolution: Resolution of the time series in hours [h]. If timeSeries is a
pandas.DataFrame() the resolution is derived from the datetime
index. optional, default: delta_T in timeSeries
:type resolution: float
:param hoursPerPeriod: Value which defines the length of a cluster period. optional, default: 24
:type hoursPerPeriod: integer
:param noTypicalPeriods: Number of typical Periods - equivalent to the number of clusters. optional, default: 10
:type noTypicalPeriods: integer
:param noSegments: Number of segments in which the typical periods shoul be subdivided - equivalent to the
number of inner-period clusters. optional, default: 10
:type noSegments: integer
:param clusterMethod: Chosen clustering method. optional, default: 'hierarchical'
|br| Options are:
* 'averaging'
* 'k_means'
* 'k_medoids'
* 'k_maxoids'
* 'hierarchical'
* 'adjacent_periods'
:type clusterMethod: string
:param evalSumPeriods: Boolean if in the clustering process also the averaged periodly values
shall be integrated additional to the periodly profiles as parameters. optional, default: False
:type evalSumPeriods: boolean
:param sameMean: Boolean which is used in the normalization procedure. If true, all time series get normalized
such that they have the same mean value. optional, default: False
:type sameMean: boolean
:param sortValues: Boolean if the clustering should be done by the periodly duration
curves (true) or the original shape of the data. optional (default: False)
:type sortValues: boolean
:param rescaleClusterPeriods: Decides if the cluster Periods shall get rescaled such that their
weighted mean value fits the mean value of the original time series. optional (default: True)
:type rescaleClusterPeriods: boolean
:param weightDict: Dictionary which weights the profiles. It is done by scaling
the time series while the normalization process. Normally all time
series have a scale from 0 to 1. By scaling them, the values get
different distances to each other and with this, they are
differently evaluated while the clustering process. optional (default: None )
:type weightDict: dict
:param extremePeriodMethod: Method how to integrate extreme Periods (peak demand, lowest temperature etc.)
into to the typical period profiles. optional, default: 'None'
|br| Options are:
* None: No integration at all.
* 'append': append typical Periods to cluster centers
* 'new_cluster_center': add the extreme period as additional cluster center. It is checked then for all
Periods if they fit better to the this new center or their original cluster center.
* 'replace_cluster_center': replaces the cluster center of the
cluster where the extreme period belongs to with the periodly profile of the extreme period. (Worst
case system design)
:type extremePeriodMethod: string
:param representationMethod: Chosen representation. If specified, the clusters are represented in the chosen
way. Otherwise, each clusterMethod has its own commonly used default representation method.
|br| Options are:
* 'meanRepresentation' (default of 'averaging' and 'k_means')
* 'medoidRepresentation' (default of 'k_medoids', 'hierarchical' and 'adjacent_periods')
* 'minmaxmeanRepresentation'
* 'durationRepresentation'/ 'distributionRepresentation'
* 'distribtionAndMinMaxRepresentation'
:type representationMethod: string
:param representationDict: Dictionary which states for each attribute whether the profiles in each cluster
should be represented by the minimum value or maximum value of each time step. This enables estimations
to the safe side. This dictionary is needed when 'minmaxmeanRepresentation' is chosen. If not specified, the
dictionary is set to containing 'mean' values only.
:type representationDict: dict
:param distributionPeriodWise: If durationRepresentation is chosen, you can choose whether the distribution of
each cluster should be separately preserved or that of the original time series only (default: True)
:type distributionPeriodWise:
:param segmentRepresentationMethod: Chosen representation for the segments. If specified, the segments are
represented in the chosen way. Otherwise, it is inherited from the representationMethod.
|br| Options are:
* 'meanRepresentation' (default of 'averaging' and 'k_means')
* 'medoidRepresentation' (default of 'k_medoids', 'hierarchical' and 'adjacent_periods')
* 'minmaxmeanRepresentation'
* 'durationRepresentation'/ 'distributionRepresentation'
* 'distribtionAndMinMaxRepresentation'
:type segmentRepresentationMethod: string
:param predefClusterOrder: Instead of aggregating a time series, a predefined grouping is taken
which is given by this list. optional (default: None)
:type predefClusterOrder: list or array
:param predefClusterCenterIndices: If predefClusterOrder is give, this list can define the representative
cluster candidates. Otherwise the medoid is taken. optional (default: None)
:type predefClusterCenterIndices: list or array
:param solver: Solver that is used for k_medoids clustering. optional (default: 'cbc' )
:type solver: string
:param roundOutput: Decimals to what the output time series get round. optional (default: None )
:type roundOutput: integer
:param addPeakMin: List of column names which's minimal value shall be added to the
typical periods. E.g.: ['Temperature']. optional, default: []
:type addPeakMin: list
:param addPeakMax: List of column names which's maximal value shall be added to the
typical periods. E.g. ['EDemand', 'HDemand']. optional, default: []
:type addPeakMax: list
:param addMeanMin: List of column names where the period with the cumulative minimal value
shall be added to the typical periods. E.g. ['Photovoltaic']. optional, default: []
:type addMeanMin: list
:param addMeanMax: List of column names where the period with the cumulative maximal value
shall be added to the typical periods. optional, default: []
:type addMeanMax: list
"""
if addMeanMin is None:
addMeanMin = []
if addMeanMax is None:
addMeanMax = []
if addPeakMax is None:
addPeakMax = []
if addPeakMin is None:
addPeakMin = []
if weightDict is None:
weightDict = {}
self.timeSeries = timeSeries
self.resolution = resolution
self.hoursPerPeriod = hoursPerPeriod
self.noTypicalPeriods = noTypicalPeriods
self.noSegments = noSegments
self.clusterMethod = clusterMethod
self.extremePeriodMethod = extremePeriodMethod
self.evalSumPeriods = evalSumPeriods
self.sortValues = sortValues
self.sameMean = sameMean
self.rescaleClusterPeriods = rescaleClusterPeriods
self.weightDict = weightDict
self.representationMethod = representationMethod
self.representationDict = representationDict
self.distributionPeriodWise = distributionPeriodWise
self.segmentRepresentationMethod = segmentRepresentationMethod
self.predefClusterOrder = predefClusterOrder
self.predefClusterCenterIndices = predefClusterCenterIndices
self.solver = solver
self.segmentation = segmentation
self.roundOutput = roundOutput
self.addPeakMin = addPeakMin
self.addPeakMax = addPeakMax
self.addMeanMin = addMeanMin
self.addMeanMax = addMeanMax
self._check_init_args()
# internal attributes
self._normalizedMean = None
return
def _check_init_args(self):
# check timeSeries and set it as pandas DataFrame
if not isinstance(self.timeSeries, pd.DataFrame):
if isinstance(self.timeSeries, dict):
self.timeSeries = pd.DataFrame(self.timeSeries)
elif isinstance(self.timeSeries, np.ndarray):
self.timeSeries = pd.DataFrame(self.timeSeries)
else:
raise ValueError(
"timeSeries has to be of type pandas.DataFrame() "
+ "or of type np.array() "
"in initialization of object of class " + type(self).__name__
)
# check if extreme periods exist in the dataframe
for peak in self.addPeakMin:
if peak not in self.timeSeries.columns:
raise ValueError(
peak
+ ' listed in "addPeakMin"'
+ " does not occur as timeSeries column"
)
for peak in self.addPeakMax:
if peak not in self.timeSeries.columns:
raise ValueError(
peak
+ ' listed in "addPeakMax"'
+ " does not occur as timeSeries column"
)
for peak in self.addMeanMin:
if peak not in self.timeSeries.columns:
raise ValueError(
peak
+ ' listed in "addMeanMin"'
+ " does not occur as timeSeries column"
)
for peak in self.addMeanMax:
if peak not in self.timeSeries.columns:
raise ValueError(
peak
+ ' listed in "addMeanMax"'
+ " does not occur as timeSeries column"
)
# derive resolution from date time index if not provided
if self.resolution is None:
try:
timedelta = self.timeSeries.index[1] - self.timeSeries.index[0]
self.resolution = float(timedelta.total_seconds()) / 3600
except AttributeError:
raise ValueError(
"'resolution' argument has to be nonnegative float or int"
+ " or the given timeseries needs a datetime index"
)
except TypeError:
try:
self.timeSeries.index = pd.to_datetime(self.timeSeries.index)
timedelta = self.timeSeries.index[1] - self.timeSeries.index[0]
self.resolution = float(timedelta.total_seconds()) / 3600
except:
raise ValueError(
"'resolution' argument has to be nonnegative float or int"
+ " or the given timeseries needs a datetime index"
)
if not (isinstance(self.resolution, int) or isinstance(self.resolution, float)):
raise ValueError("resolution has to be nonnegative float or int")
# check hoursPerPeriod
if self.hoursPerPeriod is None or self.hoursPerPeriod <= 0:
raise ValueError("hoursPerPeriod has to be nonnegative float or int")
# check typical Periods
if (
self.noTypicalPeriods is None
or self.noTypicalPeriods <= 0
or not isinstance(self.noTypicalPeriods, int)
):
raise ValueError("noTypicalPeriods has to be nonnegative integer")
self.timeStepsPerPeriod = int(self.hoursPerPeriod / self.resolution)
if not self.timeStepsPerPeriod == self.hoursPerPeriod / self.resolution:
raise ValueError(
"The combination of hoursPerPeriod and the "
+ "resulution does not result in an integer "
+ "number of time steps per period"
)
if self.segmentation:
if self.noSegments > self.timeStepsPerPeriod:
warnings.warn(
"The number of segments must be less than or equal to the number of time steps per period. "
"Segment number is decreased to number of time steps per period."
)
self.noSegments = self.timeStepsPerPeriod
# check clusterMethod
if self.clusterMethod not in self.CLUSTER_METHODS:
raise ValueError(
"clusterMethod needs to be one of "
+ "the following: "
+ "{}".format(self.CLUSTER_METHODS)
)
# check representationMethod
if (
self.representationMethod is not None
and self.representationMethod not in self.REPRESENTATION_METHODS
):
raise ValueError(
"If specified, representationMethod needs to be one of "
+ "the following: "
+ "{}".format(self.REPRESENTATION_METHODS)
)
# check representationMethod
if self.segmentRepresentationMethod is None:
self.segmentRepresentationMethod = self.representationMethod
else:
if self.segmentRepresentationMethod not in self.REPRESENTATION_METHODS:
raise ValueError(
"If specified, segmentRepresentationMethod needs to be one of "
+ "the following: "
+ "{}".format(self.REPRESENTATION_METHODS)
)
# if representationDict None, represent by maximum time steps in each cluster
if self.representationDict is None:
self.representationDict = {i: "mean" for i in list(self.timeSeries.columns)}
# sort representationDict alphabetically to make sure that the min, max or mean function is applied to the right
# column
self.representationDict = (
pd.Series(self.representationDict).sort_index(axis=0).to_dict()
)
# check extremePeriods
if self.extremePeriodMethod not in self.EXTREME_PERIOD_METHODS:
raise ValueError(
"extremePeriodMethod needs to be one of "
+ "the following: "
+ "{}".format(self.EXTREME_PERIOD_METHODS)
)
# check evalSumPeriods
if not isinstance(self.evalSumPeriods, bool):
raise ValueError("evalSumPeriods has to be boolean")
# check sortValues
if not isinstance(self.sortValues, bool):
raise ValueError("sortValues has to be boolean")
# check sameMean
if not isinstance(self.sameMean, bool):
raise ValueError("sameMean has to be boolean")
# check rescaleClusterPeriods
if not isinstance(self.rescaleClusterPeriods, bool):
raise ValueError("rescaleClusterPeriods has to be boolean")
# check predefClusterOrder
if self.predefClusterOrder is not None:
if not isinstance(self.predefClusterOrder, (list, np.ndarray)):
raise ValueError("predefClusterOrder has to be an array or list")
if self.predefClusterCenterIndices is not None:
# check predefClusterCenterIndices
if not isinstance(self.predefClusterCenterIndices, (list, np.ndarray)):
raise ValueError(
"predefClusterCenterIndices has to be an array or list"
)
elif self.predefClusterCenterIndices is not None:
raise ValueError(
'If "predefClusterCenterIndices" is defined, "predefClusterOrder" needs to be defined as well'
)
return
def _normalizeTimeSeries(self, sameMean=False):
"""
Normalizes each time series independently.
:param sameMean: Decides if the time series should have all the same mean value.
Relevant for weighting time series. optional (default: False)
:type sameMean: boolean
:returns: normalized time series
"""
min_max_scaler = preprocessing.MinMaxScaler()
normalizedTimeSeries = pd.DataFrame(
min_max_scaler.fit_transform(self.timeSeries),
columns=self.timeSeries.columns,
index=self.timeSeries.index,
)
self._normalizedMean = normalizedTimeSeries.mean()
if sameMean:
normalizedTimeSeries /= self._normalizedMean
return normalizedTimeSeries
def _unnormalizeTimeSeries(self, normalizedTimeSeries, sameMean=False):
"""
Equivalent to '_normalizeTimeSeries'. Just does the back
transformation.
:param normalizedTimeSeries: Time series which should get back transformated. required
:type normalizedTimeSeries: pandas.DataFrame()
:param sameMean: Has to have the same value as in _normalizeTimeSeries. optional (default: False)
:type sameMean: boolean
:returns: unnormalized time series
"""
from sklearn import preprocessing
min_max_scaler = preprocessing.MinMaxScaler()
min_max_scaler.fit(self.timeSeries)
if sameMean:
normalizedTimeSeries *= self._normalizedMean
unnormalizedTimeSeries = pd.DataFrame(
min_max_scaler.inverse_transform(normalizedTimeSeries),
columns=normalizedTimeSeries.columns,
index=normalizedTimeSeries.index,
)
return unnormalizedTimeSeries
def _preProcessTimeSeries(self):
"""
Normalize the time series, weight them based on the weight dict and
puts them into the correct matrix format.
"""
# first sort the time series in order to avoid bug mention in #18
self.timeSeries.sort_index(axis=1, inplace=True)
# convert the dataframe to floats
self.timeSeries = self.timeSeries.astype(float)
# normalize the time series and group them to periodly profiles
self.normalizedTimeSeries = self._normalizeTimeSeries(sameMean=self.sameMean)
for column in self.weightDict:
if self.weightDict[column] < MIN_WEIGHT:
print(
'weight of "'
+ str(column)
+ '" set to the minmal tolerable weighting'
)
self.weightDict[column] = MIN_WEIGHT
self.normalizedTimeSeries[column] = (
self.normalizedTimeSeries[column] * self.weightDict[column]
)
self.normalizedPeriodlyProfiles, self.timeIndex = unstackToPeriods(
self.normalizedTimeSeries, self.timeStepsPerPeriod
)
# check if no NaN is in the resulting profiles
if self.normalizedPeriodlyProfiles.isnull().values.any():
raise ValueError(
"Pre processed data includes NaN. Please check the timeSeries input data."
)
def _postProcessTimeSeries(self, normalizedTimeSeries, applyWeighting=True):
"""
Neutralizes the weighting the time series back and unnormalizes them.
"""
if applyWeighting:
for column in self.weightDict:
normalizedTimeSeries[column] = (
normalizedTimeSeries[column] / self.weightDict[column]
)
unnormalizedTimeSeries = self._unnormalizeTimeSeries(
normalizedTimeSeries, sameMean=self.sameMean
)
if self.roundOutput is not None:
unnormalizedTimeSeries = unnormalizedTimeSeries.round(
decimals=self.roundOutput
)
return unnormalizedTimeSeries
def _addExtremePeriods(
self,
groupedSeries,
clusterCenters,
clusterOrder,
extremePeriodMethod="new_cluster_center",
addPeakMin=None,
addPeakMax=None,
addMeanMin=None,
addMeanMax=None,
):
"""
Adds different extreme periods based on the to the clustered data,
decribed by the clusterCenters and clusterOrder.
:param groupedSeries: periodly grouped groupedSeries on which basis it should be decided,
which period is an extreme period. required
:type groupedSeries: pandas.DataFrame()
:param clusterCenters: Output from clustering with sklearn. required
:type clusterCenters: dict
:param clusterOrder: Output from clsutering with sklearn. required
:type clusterOrder: dict
:param extremePeriodMethod: Chosen extremePeriodMethod. The method. optional(default: 'new_cluster_center' )
:type extremePeriodMethod: string
:returns: - **newClusterCenters** -- The new cluster centers extended with the extreme periods.
- **newClusterOrder** -- The new cluster order including the extreme periods.
- **extremeClusterIdx** -- A list of indices where in the newClusterCenters are the extreme
periods located.
"""
# init required dicts and lists
self.extremePeriods = {}
extremePeriodNo = []
ccList = [center.tolist() for center in clusterCenters]
# check which extreme periods exist in the profile and add them to
# self.extremePeriods dict
for column in self.timeSeries.columns:
if column in addPeakMax:
stepNo = groupedSeries[column].max(axis=1).idxmax()
# add only if stepNo is not already in extremePeriods
# if it is not already a cluster center
if (
stepNo not in extremePeriodNo
and groupedSeries.loc[stepNo, :].values.tolist() not in ccList
):
max_col = self._append_col_with(column, " max.")
self.extremePeriods[max_col] = {
"stepNo": stepNo,
"profile": groupedSeries.loc[stepNo, :].values,
"column": column,
}
extremePeriodNo.append(stepNo)
if column in addPeakMin:
stepNo = groupedSeries[column].min(axis=1).idxmin()
# add only if stepNo is not already in extremePeriods
# if it is not already a cluster center
if (
stepNo not in extremePeriodNo
and groupedSeries.loc[stepNo, :].values.tolist() not in ccList
):
min_col = self._append_col_with(column, " min.")
self.extremePeriods[min_col] = {
"stepNo": stepNo,
"profile": groupedSeries.loc[stepNo, :].values,
"column": column,
}
extremePeriodNo.append(stepNo)
if column in addMeanMax:
stepNo = groupedSeries[column].mean(axis=1).idxmax()
# add only if stepNo is not already in extremePeriods
# if it is not already a cluster center
if (
stepNo not in extremePeriodNo
and groupedSeries.loc[stepNo, :].values.tolist() not in ccList
):
mean_max_col = self._append_col_with(column, " daily max.")
self.extremePeriods[mean_max_col] = {
"stepNo": stepNo,
"profile": groupedSeries.loc[stepNo, :].values,
"column": column,
}
extremePeriodNo.append(stepNo)
if column in addMeanMin:
stepNo = groupedSeries[column].mean(axis=1).idxmin()
# add only if stepNo is not already in extremePeriods and
# if it is not already a cluster center
if (
stepNo not in extremePeriodNo
and groupedSeries.loc[stepNo, :].values.tolist() not in ccList
):
mean_min_col = self._append_col_with(column, " daily min.")
self.extremePeriods[mean_min_col] = {
"stepNo": stepNo,
"profile": groupedSeries.loc[stepNo, :].values,
"column": column,
}
extremePeriodNo.append(stepNo)
for periodType in self.extremePeriods:
# get current related clusters of extreme periods
self.extremePeriods[periodType]["clusterNo"] = clusterOrder[
self.extremePeriods[periodType]["stepNo"]
]
# init new cluster structure
newClusterCenters = []
newClusterOrder = clusterOrder
extremeClusterIdx = []
# integrate extreme periods to clusters
if extremePeriodMethod == "append":
# attach extreme periods to cluster centers
for i, cluster_center in enumerate(clusterCenters):
newClusterCenters.append(cluster_center)
for i, periodType in enumerate(self.extremePeriods):
extremeClusterIdx.append(len(newClusterCenters))
newClusterCenters.append(self.extremePeriods[periodType]["profile"])
newClusterOrder[self.extremePeriods[periodType]["stepNo"]] = i + len(
clusterCenters
)
elif extremePeriodMethod == "new_cluster_center":
for i, cluster_center in enumerate(clusterCenters):
newClusterCenters.append(cluster_center)
# attach extrem periods to cluster centers and consider for all periods
# if the fit better to the cluster or the extrem period
for i, periodType in enumerate(self.extremePeriods):
extremeClusterIdx.append(len(newClusterCenters))
newClusterCenters.append(self.extremePeriods[periodType]["profile"])
self.extremePeriods[periodType]["newClusterNo"] = i + len(
clusterCenters
)
for i, cPeriod in enumerate(newClusterOrder):
# caclulate euclidean distance to cluster center
cluster_dist = sum(
(groupedSeries.iloc[i].values - clusterCenters[cPeriod]) ** 2
)
for ii, extremPeriodType in enumerate(self.extremePeriods):
# exclude other extreme periods from adding to the new
# cluster center
isOtherExtreme = False
for otherExPeriod in self.extremePeriods:
if (
i == self.extremePeriods[otherExPeriod]["stepNo"]
and otherExPeriod != extremPeriodType
):
isOtherExtreme = True
# calculate distance to extreme periods
extperiod_dist = sum(
(
groupedSeries.iloc[i].values
- self.extremePeriods[extremPeriodType]["profile"]
)
** 2
)
# choose new cluster relation
if extperiod_dist < cluster_dist and not isOtherExtreme:
newClusterOrder[i] = self.extremePeriods[extremPeriodType][
"newClusterNo"
]
elif extremePeriodMethod == "replace_cluster_center":
# Worst Case Clusterperiods
newClusterCenters = clusterCenters
for periodType in self.extremePeriods:
index = groupedSeries.columns.get_loc(
self.extremePeriods[periodType]["column"]
)
newClusterCenters[self.extremePeriods[periodType]["clusterNo"]][
index
] = self.extremePeriods[periodType]["profile"][index]
if (
not self.extremePeriods[periodType]["clusterNo"]
in extremeClusterIdx
):
extremeClusterIdx.append(
self.extremePeriods[periodType]["clusterNo"]
)
return newClusterCenters, newClusterOrder, extremeClusterIdx
def _append_col_with(self, column, append_with=" max."):
"""Appends a string to the column name. For MultiIndexes, which turn out to be
tuples when this method is called, only last level is changed"""
if isinstance(column, str):
return column + append_with
elif isinstance(column, tuple):
col = list(column)
col[-1] = col[-1] + append_with
return tuple(col)
def _rescaleClusterPeriods(self, clusterOrder, clusterPeriods, extremeClusterIdx):
"""
Rescale the values of the clustered Periods such that mean of each time
series in the typical Periods fits the mean value of the original time
series, without changing the values of the extremePeriods.
"""
weightingVec = pd.Series(self._clusterPeriodNoOccur).values
typicalPeriods = pd.DataFrame(
clusterPeriods, columns=self.normalizedPeriodlyProfiles.columns
)
idx_wo_peak = np.delete(typicalPeriods.index, extremeClusterIdx)
for column in self.timeSeries.columns:
diff = 1
sum_raw = self.normalizedPeriodlyProfiles[column].sum().sum()
sum_peak = sum(
weightingVec[extremeClusterIdx]
* typicalPeriods[column].loc[extremeClusterIdx, :].sum(axis=1)
)
sum_clu_wo_peak = sum(
weightingVec[idx_wo_peak]
* typicalPeriods[column].loc[idx_wo_peak, :].sum(axis=1)
)
# define the upper scale dependent on the weighting of the series
scale_ub = 1.0
if self.sameMean:
scale_ub = (
scale_ub
* self.timeSeries[column].max()
/ self.timeSeries[column].mean()
)
if column in self.weightDict:
scale_ub = scale_ub * self.weightDict[column]
# difference between predicted and original sum
diff = abs(sum_raw - (sum_clu_wo_peak + sum_peak))
# use while loop to rescale cluster periods
a = 0
while diff > sum_raw * TOLERANCE and a < MAX_ITERATOR:
# rescale values
typicalPeriods.loc[idx_wo_peak, column] = (
typicalPeriods[column].loc[idx_wo_peak, :].values
* (sum_raw - sum_peak)
/ sum_clu_wo_peak
)
# reset values higher than the upper sacle or less than zero
typicalPeriods[column][typicalPeriods[column] > scale_ub] = scale_ub
typicalPeriods[column][typicalPeriods[column] < 0.0] = 0.0
typicalPeriods[column] = typicalPeriods[column].fillna(0.0)
# calc new sum and new diff to orig data
sum_clu_wo_peak = sum(
weightingVec[idx_wo_peak]
* typicalPeriods[column].loc[idx_wo_peak, :].sum(axis=1)
)
diff = abs(sum_raw - (sum_clu_wo_peak + sum_peak))
a += 1
if a == MAX_ITERATOR:
deviation = str(round((diff / sum_raw) * 100, 2))
warnings.warn(
'Max iteration number reached for "'
+ str(column)
+ '" while rescaling the cluster periods.'
+ " The integral of the aggregated time series deviates by: "
+ deviation
+ "%"
)
return typicalPeriods.values
def _clusterSortedPeriods(self, candidates, n_init=20):
"""
Runs the clustering algorithms for the sorted profiles within the period
instead of the original profiles. (Duration curve clustering)
"""
# initialize
normalizedSortedPeriodlyProfiles = copy.deepcopy(
self.normalizedPeriodlyProfiles
)
for column in self.timeSeries.columns:
# sort each period individually
df = normalizedSortedPeriodlyProfiles[column]
values = df.values
values.sort(axis=1)
values = values[:, ::-1]
normalizedSortedPeriodlyProfiles[column] = pd.DataFrame(
values, df.index, df.columns
)
sortedClusterValues = normalizedSortedPeriodlyProfiles.values
(
altClusterCenters,
self.clusterCenterIndices,
clusterOrders_C,
) = aggregatePeriods(
sortedClusterValues,
n_clusters=self.noTypicalPeriods,
n_iter=30,
solver=self.solver,
clusterMethod=self.clusterMethod,
representationMethod=self.representationMethod,
representationDict=self.representationDict,
distributionPeriodWise=self.distributionPeriodWise,
timeStepsPerPeriod=self.timeStepsPerPeriod,
)
clusterCenters_C = []
# take the clusters and determine the most representative sorted
# period as cluster center
for clusterNum in np.unique(clusterOrders_C):
indice = np.where(clusterOrders_C == clusterNum)[0]
if len(indice) > 1:
# mean value for each time step for each time series over
# all Periods in the cluster
currentMean_C = sortedClusterValues[indice].mean(axis=0)
# index of the period with the lowest distance to the cluster
# center
mindistIdx_C = np.argmin(
np.square(sortedClusterValues[indice] - currentMean_C).sum(axis=1)
)
# append original time series of this period
medoid_C = candidates[indice][mindistIdx_C]
# append to cluster center
clusterCenters_C.append(medoid_C)
else:
# if only on period is part of the cluster, add this index
clusterCenters_C.append(candidates[indice][0])
return clusterCenters_C, clusterOrders_C
[docs] def createTypicalPeriods(self):
"""
Clusters the Periods.
:returns: **self.typicalPeriods** -- All typical Periods in scaled form.
"""
self._preProcessTimeSeries()
# check for additional cluster parameters
if self.evalSumPeriods:
evaluationValues = (
self.normalizedPeriodlyProfiles.stack(level=0)
.sum(axis=1)
.unstack(level=1)
)
# how many values have to get deleted later
delClusterParams = -len(evaluationValues.columns)
candidates = np.concatenate(
(self.normalizedPeriodlyProfiles.values, evaluationValues.values),
axis=1,
)
else:
delClusterParams = None
candidates = self.normalizedPeriodlyProfiles.values
# skip aggregation procedure for the case of a predefined cluster sequence and get only the correct representation
if not self.predefClusterOrder is None:
self._clusterOrder = self.predefClusterOrder
# check if representatives are defined
if not self.predefClusterCenterIndices is None:
self.clusterCenterIndices = self.predefClusterCenterIndices
self.clusterCenters = candidates[self.predefClusterCenterIndices]
else:
# otherwise take the medoids
self.clusterCenters, self.clusterCenterIndices = representations(
candidates,
self._clusterOrder,
default="medoidRepresentation",
representationMethod=self.representationMethod,
representationDict=self.representationDict,
timeStepsPerPeriod=self.timeStepsPerPeriod,
)
else:
cluster_duration = time.time()
if not self.sortValues:
# cluster the data
(
self.clusterCenters,
self.clusterCenterIndices,
self._clusterOrder,
) = aggregatePeriods(
candidates,
n_clusters=self.noTypicalPeriods,
n_iter=100,
solver=self.solver,
clusterMethod=self.clusterMethod,
representationMethod=self.representationMethod,
representationDict=self.representationDict,
distributionPeriodWise=self.distributionPeriodWise,
timeStepsPerPeriod=self.timeStepsPerPeriod,
)
else:
self.clusterCenters, self._clusterOrder = self._clusterSortedPeriods(
candidates
)
self.clusteringDuration = time.time() - cluster_duration
# get cluster centers without additional evaluation values
self.clusterPeriods = []
for i, cluster_center in enumerate(self.clusterCenters):
self.clusterPeriods.append(cluster_center[:delClusterParams])
if not self.extremePeriodMethod == "None":
# overwrite clusterPeriods and clusterOrder
(
self.clusterPeriods,
self._clusterOrder,
self.extremeClusterIdx,
) = self._addExtremePeriods(
self.normalizedPeriodlyProfiles,
self.clusterPeriods,
self._clusterOrder,
extremePeriodMethod=self.extremePeriodMethod,
addPeakMin=self.addPeakMin,
addPeakMax=self.addPeakMax,
addMeanMin=self.addMeanMin,
addMeanMax=self.addMeanMax,
)
else:
self.extremeClusterIdx = []
# get number of appearance of the the typical periods
nums, counts = np.unique(self._clusterOrder, return_counts=True)
self._clusterPeriodNoOccur = {num: counts[ii] for ii, num in enumerate(nums)}
if self.rescaleClusterPeriods:
self.clusterPeriods = self._rescaleClusterPeriods(
self._clusterOrder, self.clusterPeriods, self.extremeClusterIdx
)
# if additional time steps have been added, reduce the number of occurrence of the typical period
# which is related to these time steps
if not len(self.timeSeries) % self.timeStepsPerPeriod == 0:
self._clusterPeriodNoOccur[self._clusterOrder[-1]] -= (
1
- float(len(self.timeSeries) % self.timeStepsPerPeriod)
/ self.timeStepsPerPeriod
)
# put the clustered data in pandas format and scale back
self.normalizedTypicalPeriods = pd.DataFrame(
self.clusterPeriods, columns=self.normalizedPeriodlyProfiles.columns
).stack(level="TimeStep")
if self.segmentation:
from tsam.utils.segmentation import segmentation
(
self.segmentedNormalizedTypicalPeriods,
self.predictedSegmentedNormalizedTypicalPeriods,
) = segmentation(
self.normalizedTypicalPeriods,
self.noSegments,
self.timeStepsPerPeriod,
representationMethod=self.segmentRepresentationMethod,
representationDict=self.representationDict,
distributionPeriodWise=self.distributionPeriodWise,
)
self.normalizedTypicalPeriods = (
self.segmentedNormalizedTypicalPeriods.reset_index(level=3, drop=True)
)
self.typicalPeriods = self._postProcessTimeSeries(self.normalizedTypicalPeriods)
# check if original time series boundaries are not exceeded
if np.array(
self.typicalPeriods.max(axis=0) > self.timeSeries.max(axis=0)
).any():
warning_list = self.typicalPeriods.max(axis=0) < self.timeSeries.max(axis=0)
warnings.warn(
"Something went wrong... At least one maximal value of the " +
"aggregated time series exceeds the maximal value " +
"the input time series for: " +
"{}".format(list(warning_list[warning_list>0].index))
)
if np.array(
self.typicalPeriods.min(axis=0) < self.timeSeries.min(axis=0)
).any():
warning_list = self.typicalPeriods.min(axis=0) < self.timeSeries.min(axis=0)
warnings.warn(
"Something went wrong... At least one minimal value of the " +
"aggregated time series exceeds the minimal value " +
"the input time series for: " +
"{}".format(list(warning_list[warning_list>0].index))
)
return self.typicalPeriods
@property
def stepIdx(self):
"""
Index inside a single cluster
"""
if self.segmentation:
return [ix for ix in range(0, self.noSegments)]
else:
return [ix for ix in range(0, self.timeStepsPerPeriod)]
@property
def clusterPeriodIdx(self):
"""
Index of the clustered periods
"""
if not hasattr(self, "clusterOrder"):
self.createTypicalPeriods()
return np.sort(np.unique(self._clusterOrder))
@property
def clusterOrder(self):
"""
The sequence/order of the typical period to represent
the original time series
"""
if not hasattr(self, "_clusterOrder"):
self.createTypicalPeriods()
return self._clusterOrder
@property
def clusterPeriodNoOccur(self):
"""
How often does a typical period occur in the original time series
"""
if not hasattr(self, "clusterOrder"):
self.createTypicalPeriods()
return self._clusterPeriodNoOccur
@property
def clusterPeriodDict(self):
"""
Time series data for each period index as dictionary
"""
if not hasattr(self, "_clusterOrder"):
self.createTypicalPeriods()
if not hasattr(self, "_clusterPeriodDict"):
self._clusterPeriodDict = {}
for column in self.typicalPeriods:
self._clusterPeriodDict[column] = self.typicalPeriods[column].to_dict()
return self._clusterPeriodDict
@property
def segmentDurationDict(self):
"""
Segment duration in time steps for each period index as dictionary
"""
if not hasattr(self, "_clusterOrder"):
self.createTypicalPeriods()
if not hasattr(self, "_segmentDurationDict"):
if self.segmentation:
self._segmentDurationDict = (
self.segmentedNormalizedTypicalPeriods.drop(
self.segmentedNormalizedTypicalPeriods.columns, axis=1
)
.reset_index(level=3, drop=True)
.reset_index(2)
.to_dict()
)
else:
self._segmentDurationDict = self.typicalPeriods.drop(
self.typicalPeriods.columns, axis=1
)
self._segmentDurationDict["Segment Duration"] = 1
self._segmentDurationDict = self._segmentDurationDict.to_dict()
warnings.warn(
"Segmentation is turned off. All segments are consistent the time steps."
)
return self._segmentDurationDict
[docs] def predictOriginalData(self):
"""
Predicts the overall time series if every period would be placed in the
related cluster center
:returns: **predictedData** (pandas.DataFrame) -- DataFrame which has the same shape as the original one.
"""
if not hasattr(self, "_clusterOrder"):
self.createTypicalPeriods()
# list up typical periods according to their order of occurrence using the _clusterOrder.
new_data = []
for label in self._clusterOrder:
# if segmentation is used, use the segmented typical periods with predicted time steps with the same number
# of time steps as unsegmented typical periods
if self.segmentation:
new_data.append(
self.predictedSegmentedNormalizedTypicalPeriods.loc[label, :]
.unstack()
.values
)
else:
# new_data.append(self.clusterPeriods[label])
new_data.append(
self.normalizedTypicalPeriods.loc[label, :].unstack().values
)
# back in matrix
clustered_data_df = pd.DataFrame(
new_data,
columns=self.normalizedPeriodlyProfiles.columns,
index=self.normalizedPeriodlyProfiles.index,
)
clustered_data_df = clustered_data_df.stack(level="TimeStep")
# back in form
self.normalizedPredictedData = pd.DataFrame(
clustered_data_df.values[: len(self.timeSeries)],
index=self.timeSeries.index,
columns=self.timeSeries.columns,
)
# normalize again if sameMean = True to avoid doubled unnormalization when using _postProcessTimeSeries after
# createTypicalPeriods has been called
if self.sameMean:
self.normalizedPredictedData /= self._normalizedMean
self.predictedData = self._postProcessTimeSeries(
self.normalizedPredictedData, applyWeighting=False
)
return self.predictedData
[docs] def indexMatching(self):
"""
Relates the index of the original time series with the indices
represented by the clusters
:returns: **timeStepMatching** (pandas.DataFrame) -- DataFrame which has the same shape as the original one.
"""
if not hasattr(self, "_clusterOrder"):
self.createTypicalPeriods()
# create aggregated period and time step index lists
periodIndex = []
stepIndex = []
for label in self._clusterOrder:
for step in range(self.timeStepsPerPeriod):
periodIndex.append(label)
stepIndex.append(step)
# create a dataframe
timeStepMatching = pd.DataFrame(
[periodIndex, stepIndex],
index=["PeriodNum", "TimeStep"],
columns=self.timeIndex,
).T
# if segmentation is chosen, append another column stating which
if self.segmentation:
segmentIndex = []
for label in self._clusterOrder:
segmentIndex.extend(
np.repeat(
self.segmentedNormalizedTypicalPeriods.loc[
label, :
].index.get_level_values(0),
self.segmentedNormalizedTypicalPeriods.loc[
label, :
].index.get_level_values(1),
).values
)
timeStepMatching = pd.DataFrame(
[periodIndex, stepIndex, segmentIndex],
index=["PeriodNum", "TimeStep", "SegmentIndex"],
columns=self.timeIndex,
).T
return timeStepMatching
[docs] def accuracyIndicators(self):
"""
Compares the predicted data with the original time series.
:returns: **pd.DataFrame(indicatorRaw)** (pandas.DataFrame) -- Dataframe containing indicators evaluating the
accuracy of the
aggregation
"""
if not hasattr(self, "predictedData"):
self.predictOriginalData()
indicatorRaw = {
"RMSE": {},
"RMSE_duration": {},
"MAE": {},
} # 'Silhouette score':{},
for column in self.normalizedTimeSeries.columns:
if self.weightDict:
origTS = self.normalizedTimeSeries[column] / self.weightDict[column]
else:
origTS = self.normalizedTimeSeries[column]
predTS = self.normalizedPredictedData[column]
indicatorRaw["RMSE"][column] = np.sqrt(mean_squared_error(origTS, predTS))
indicatorRaw["RMSE_duration"][column] = np.sqrt(
mean_squared_error(
origTS.sort_values(ascending=False).reset_index(drop=True),
predTS.sort_values(ascending=False).reset_index(drop=True),
)
)
indicatorRaw["MAE"][column] = mean_absolute_error(origTS, predTS)
return pd.DataFrame(indicatorRaw)
[docs] def totalAccuracyIndicators(self):
"""
Derives the accuracy indicators over all time series
"""
return np.sqrt(self.accuracyIndicators().pow(2).sum()/len(self.normalizedTimeSeries.columns))
