HAL Id: tel-01557762
https://tel.archives-ouvertes.fr/tel-01557762
Submitted on 6 Jul 2017
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Ni Ding
To cite this version:
Spé ialité: Génie Éle trique
Arrêtéministériel: 7 Aut 2006
Présentée par
Ni DING
Thèse dirigée par Yvon BÉSANGERet
odirigée par Frédéri WURTZ
préparée ausein du
Laboratoire G2ELAB
dans l'É ole Do torale: EEATS
Load models for operation and planning of
ele tri ity distribution networks with smart
metering data
Thèse soutenue publiquementle 30 Novembre 2012,
devant lejury omposé de:
Pr. Nouredine Hadjsaid
GrenobleINP, Président
Pr. Carlo Alberto Nu i
UniversitédeBologne,Rapporteur
Pr. Corinne Alonso
UniversitédeToulouse,Rapporteur
Pr. Didier Mayer
MinedeParis,Membre
Pr. Yvon Bésanger
GrenobleINP,Membre
Dr. Frédéri Wurtz
CNRSGrenoble,Membre
Invités:
M. Olivier Devaux
EDFR&D
M. Alain Glatigny
A knowledgments xiii
Notations xix
1 General introdu tion: the new problemati of load models
in the smart grid ontext 1
1.1 Ba kground: smartgrid and smart meters for load modeling . . 2
1.2 Motivation and obje tives . . . 3
1.2.a Fornetworkoperationneed. . . 4
1.2.b Fornetworkplanningneed . . . 5
1.3 Contributions of the thesis . . . 6
1.4 S ope and organization of the dissertation . . . 7
A Short-term load fore asting models for monitoring and state es-timator 11 2 Load fore astingte hniques andshort-term modelframework 13 2.1 Literature review . . . 14
2.1.a Fore astingleadtimesandinuen e fa tors. . . 14
2.1.b Fore astingmethods . . . 16
2.1.b-i Classi alapproa h . . . 18
2.1.b-ii Arti ialintelligentapproa h . . . 25
2.1.b-iii Hybridmodels . . . 35
2.1. Literaturereview on lusionsandperspe tives . . . 37
2.2 Data des ription . . . 40
2.2.a MV/LVsubstation . . . 40
2.2.a-i Temperatureinuen e . . . 41
2.2.a-ii Daytypeinuen e . . . 42
2.2.a-iii Timeinuen e . . . 42
2.2.b MVfeeder . . . 45
2.3 Choi es of Time series and NN methods . . . 45
2.4 Performan e riteria and referen e ase . . . 46
2.4.a Performan e riteria: MAPEandMAE . . . 46
2.4.b Referen e ase: thenaivemodel . . . 47
3 Time series model 49
3.1 Additive time series model and pro edure overview . . . 50
3.2 Statisti al tools . . . 51
3.2.a DummyVariableRegression . . . 51
3.2.b TrendComponentEstimation . . . 52
3.2. Cy li ComponentEstimation . . . 52
3.2.d TestsofStationarity . . . 53
3.2.e SmoothedPeriodogram . . . 53
3.2.f RegressionModelwithFourierComponents . . . 54
3.2.g ANOVA NullityTest . . . 54
3.2.h CompleteFore astingModel . . . 55
3.3 Appli ation example results . . . 55
3.3.a Trainingset . . . 55
3.3.b Testset . . . 57
3.3. ResidualAnalysis . . . 60
3.3. -i Normality. . . 60
3.3. -ii Independen e . . . 61
3.4 Weather un ertainty . . . 62
3.5 Con lusion . . . 64
4 Neural network model 67 4.1 Ma hine learning te hnique . . . 68
4.2 Multi Layer Per eptrons and trainingpro ess . . . 69
4.3 Model design. . . 71
4.3.a Variablesele tion . . . 71
4.3.b Modelsele tion . . . 73
4.3.b-i Modelsele tionmethodology . . . 73
4.3.b-ii Assessmentofthegeneralizationabilityofthemodels . . . 74
4.4 Numeri al illustration . . . 76
4.4.a Framework . . . 76
4.4.b Modeldesign: anillustrativeexample . . . 77
4.4.b-i Variablesele tionexample. . . 77
4.4.b-ii Sele tingthebest modelforagiven omplexity . . . 81
4.4.b-iii Complexitysele tionexample. . . 81
4.4. Results . . . 83
4.5 Overall omparison with the time series model . . . 85
4.6 Con lusion and perspe tive . . . 86
B Load estimation models for distribution network planning 89 5 Loadresear hproje tsindistributionnetworks: stateofthe art 91 5.1 De ision making in distribution network planning . . . 92
5.1.b Typi alLoadProle(TLP) . . . 95
5.2 Load resear h proje ts in different ountries . . . 96
5.2.a FinlandDSOmodel . . . 97
5.2.b DenmarkDongEnergy . . . 98
5.2. NorwaySINTEFEnergyResear h . . . 99
5.2.d Taipowersystem. . . 99
5.3 Fren h load resear h proje t . . . .100
5.3.a Datades ription. . . .102
5.3.b EDFBAGHEERA model. . . .103
5.3.b-i TMB temperatureandbasi model. . . .104
5.3.b-ii Common oe ientestimation . . . .105
5.3.b-iii Spe i parameterestimation . . . .105
5.3.b-iv Illustrativeexampleandmodel's output . . . .107
5.4 Con lusion . . . .111
6 Nonparametri model 113 6.1 Nonparametri model. . . .115
6.1.a Statisti altests . . . .116
6.1.b Kerneldensityestimation. . . .117
6.1. CUSUMalgorithm . . . .117
6.1.d Kernelregression . . . .118
6.1.e Smoothingparametersele tion: ross-validationte hnique. . . .119
6.2 Computational example . . . .120
6.2.a Illustrativeexampleresults . . . .121
6.2.b ComparisonwiththeBAGHEERAmodel. . . .124
6.3 Validation study . . . .127
6.4 Dis ussion . . . .129
6.4.a Citationsoftheupper-bounddenitionsin EDFreports. . . .130
6.4.b Upperboundin thenonparametri models . . . .131
6.4. Validationtrialontheupper-bound estimation . . . .132
6.5 Con lusion and perspe tive . . . .137
7 General on lusion and perspe tive 139 7.1 Con lusion . . . .139
7.2 Perspe tive . . . .140
Bibliographie 154
Appendi es 155
A Time series model's result summary 155
B Binary hypothesis test 157
D Comparisonresultsof naivemodel, timeseriesmodeland
neu-ralnetwork model 161
E Résumé français 169
E.1 Introdu tion générale: la nouvelle problématique du modèle
de harge dans le ontexte du réseau intelligent . . . .169
E.1.a Réseauintelligentet ompteurs intelligentspourlesmodèlesde harge . . .169
E.1.b Obje tifset plandurésuméfrançais . . . .170
E.1. Contributiondethèse . . . .172
E.2 Modèle de harge prédi tif ourt terme pour la onduite et l'estimateur d'état. . . .173
E.2.a Méthodesdelaprévisionde hargedanslalittérature . . . .174
E.2.b Des riptiondedonnées . . . .178
E.2. Choixdesméthodes: série hronologiqueet réseaudeneurones . . . .180
E.2.d Critèresdeperforman eet modèlederéféren e. . . .181
E.2.e Modèlesérie hronologique . . . .182
E.2.f Modèleréseaudeneurones . . . .187
E.2.f-i Con eptiondumodèle . . . .188
E.2.f-ii Comparaisonglobaleave lemodèledesérie hronologique. . . .191
E.3 Modèled'estimation de hargepour laplanifi ation du réseau de distribution . . . .193
E.3.a ModèleBAGHEERA . . . .195
E.3.b Modèlenonparamétrique. . . .198
1.1 Availablemeasurements intheFren h distribution networks . . . 3
1.2 Relationshipamongfore asting models,SE, andADA fun tions. . . 5
2.1 Summaryof loadfore asting methods intwo dimensions . . . 17
2.2 Singleper eptronstru ture . . . 27
2.3 One-hidden-layer network stru ture . . . 27
2.4 Supervised learningpro edure . . . 28
2.5 Re urrent neural network stru ture . . . 29
2.6 FuzzyLogi pro ess . . . 34
2.7 Fuzzylogi : inputvariables membership fun tion . . . 35
2.8 Fuzzylogi : outputvariablesmembership fun tion. . . 35
2.9 Daily average load and temperature data through
414
days (from Sept. 9, 2009 to O t. 27,2010)ofsubstation CE_MOU (mainly residential) . . . 402.10 Dailyaverage loadthrough
414
days (from Sept. 9,2009 to O t. 27,2010) ofsubstation VI_LOG (mixedservi ese tor andindustrial) . . . 412.11 Dailyaverage loadthrough
414
days (from Sept. 9,2009 to O t. 27,2010) ofsubstation CE_FRO (an industrial lient) . . . 412.12 Normal week ompared to the week with a national holiday of Substation CE_FRO (an industrial lient) . . . 43
2.13 Similarity index al ulatedbased onall days of substationCE_MOU . . . . 44
2.14 Similarityindexwithoutweekendsandpubli holidaysofsubstationCE_MOU 44 2.15 MVfeeders and position of onne ted MV/LVsubstations . . . 45
3.1 Stepsof the designed time seriesfore asting method . . . 51
3.2 Trainingset andtest setperiodsof theavailable data. . . 55
3.3 A weekly onsumption pattern (O tober 5,2009 to O tober11, 2009) of a mixedindustrialand servi ese torsubstation VI_LOG. . . 56
3.4 SubstationVI_LOG,MAE riteria al ulatedonthetrainingset(117days) fordierent slidingwindow sizes(weeks). . . 56
3.5 PeriodogramofthedetrendedtrainingdatasetsmoothedbytheDaniellkernel 57 3.6 Substation VI_LOG, omparison of the fore asting results with the real measurementson the test setperiod(297 days). . . 58
3.7 Substation VI_LOG,two-day-ahead load fore astingresults on weekdays . . 58
3.8 Substation VI_LOG,two-day-ahead load fore astingresults on weekends . . 59
3.9 Substation VI_LOG,densityfun tion plotand umulative densityfun tion plotof the residual. . . 61
3.11 Histogram of the Gaussian distributed temperature un ertainty adding to
thea tual temperature . . . 63
3.12 Three-day fore asting temperatures ompared to thea tual temperatures . . 64
4.1 Orthogonal forward rankingpro ess. . . 73
4.2 Neural network sele tion pro edure . . . 74
4.3 Separation of the load urve into the daily averagepowerand theintraday power variation . . . 77
4.4 Generation of se ondary variables and probe variables. . . 78
4.5 Cumulative probability for a probe variable to have a better rank than a andidate variable. . . 80
4.6 Modelsele tion for theintradaypowervariationmodel . . . 82
4.7 Neural network omplexitysele tion strategieswithVLOOs oreand lever-age distribution . . . 82
5.1 Network de ision makingpro edure . . . 93
5.2 Example of oin iden e fa tor al ulation . . . 94
5.3 Distribution Load Estimation (DLE)pro ess . . . 97
5.4 Voltage-drop andtap hanger adjustment. . . .101
5.5 Two-year(July 01, 2004
∼
June 30, 2006) daily average loads of o-peak/ on-peak option lient no.5 . . . .1035.6 Two-year(July01,2004
∼
June 30,2006)dailyaverageloadsofbasi option lient no.18 . . . .1035.7 O-peak/on-peak option lientno.5: urve tting on o-peak dailyenergy use . . . .108
5.8 O-peak/on-peak option lientno.5: urve tting on on-peak dailyenergy use . . . .108
5.9 Basi option lientno.18: urve tting ondailyenergy use. . . .109
5.10 O-peak/on-peak option lient no.5: outputs of the BAGHEERA model, TMB loadestimations on weekdays . . . .110
5.11 O-peak/ on-peak option lient no.5 : omparison of TMB weekend's and weekday'sloadestimation. . . .111
6.1 Overviewofthe nonparametri model . . . .116
6.2 Statisti al testspro edure. . . .117
6.3 Data diagram: histori al data,1st-yeardata, and2nd-yeardata. . . .120
6.4 O-peak/on-peak option lientno.5: statisti altestsresultof thermosensi-tive he k . . . .121
6.5 O-peak/on-peak option lient no.5: CUSUM hartofdaily average power .121 6.6 O-peak/on-peak option lient no.5: separation result of one year's power databyCUSUMalgorithm . . . .122
6.7 O-peak/on-peak option lient no.5: weekdayminimumpower estimations .122 6.8 O-peak/on-peak option lient no.5: statisti al tests result for the data oheren e he k . . . .123
6.10 NW,LL,and LL2 regressors,indi ating therelationshipbetween the
varia-tionof temperature and the lient's dailypower onsumption . . . .124
6.11 O-peak/on-peakoption lientno.5: presentation ofun ertainty ofa sample125
6.12 O-peak/on-peak option lient no.5: maximumpowerestimation of
week-dayloads. . . .125
6.13 SumSquareErrors(SSE)softheBAGHEERAestimator,NW,LL,andLL2
estimatorson the test data . . . .126
6.14 Study ases and s enariosin thevalidation study. . . .127
6.15 Study ase no.1, s enario1: o-peak/on-peakoption lients, omparison of
SSEs of BAGHEERA, NW, LL and LL2 estimators on the days below 0
degreeduring the se ondyear . . . .128
6.16 Study ase no.1, s enario2: o-peak/on-peakoption lients, omparison of
SSEsofBAGHEERA,NW,LLandLL2estimators ontheo-peakhoursof
the days below0 degreeduring these ondyear . . . .128
6.17 Study ase no.2, s enario1: o-peak/on-peakoption lients, omparison of
SSEs of BAGHEERA, NW, LL and LL2 estimators on the 30 oldestdays
ofthe se ond-yeardata . . . .129
6.18 Study ase no.2, s enario2: o-peak/on-peakoption lients, omparison of
SSEsofBAGHEERA,NW,LLandLL2estimators ontheo-peakhoursof
the 30 oldest days ofthese ond-yeardata . . . .129
6.19 10%hourly power ex ess probability threshold and median value for every
time step. . . .132
6.20 Summary of the upper-bound omparison of the real measurements, the
BAGHEERA model,and nonparametri models . . . .133
6.21 Power onsumption of lient no.22 during two years (July01, 2004
∼
June 30,2006) . . . .1346.22 Power onsumption of lient no.17 during two years (July01, 2004
∼
June 30,2006) . . . .1356.23 30-minute timestep standarddeviation(sd) of lientNo.17 . . . .135
E.1 Relation entre les modèles de harge prédi tifs, l'estimateur d'état, et les
fon tionsavan ées du réseau . . . .171
E.2 Résumédes méthodesde hargeprédi tives endeux dimensions. . . .176
E.3 Courbede hargeettempératurejournalièrependant
414
jours(du9/9/2009 au 27/10/2010) du poste HTA/BT CE_MOU ( onne té prin ipalement àdes lientsrésidentiels) . . . .179
E.4 Courbede hargejournalièrependant
414
jours(du9/9/2009au27/10/2010) du poste HTA/BT VI_LOG ( onne té aux lients mixtes tertiaires etin-dustriels). . . .179
E.5 Courbede hargejournalièrependant
414
jours(du9/9/2009au27/10/2010) duposteHTA/BTCE_FRO ( onne té àunseul lient industriel) . . . .180E.6 Etapes pour onstruire le modèle série hronologique pour la prévision de
harge . . . .183
E.7 Pro éduredu lassement par laproje tion orthogonale deGram-S hmidt . .189
E.9 La prise desdé isions dansleréseau de distribution . . . .194
E.10Client no.5 option heure reuse/pleine: ajustement de ourbe sur l'énergie
journalièrependantlesheurespleines. L'indi eHPsignieHeurePleine
etl'indi e HCsignie HeureCreuse. . . .196
E.11Client no.18 option de base: ajustement de ourbe surl'énergiejournalière. 196
E.12Clientno.5 option heure reuse/pleine: TMB estimationsde la harge
pen-dant lesjours ouvrables . . . .197
2.1 Dierent time horizon loadfore asts . . . 15
2.2 Summaryof loadfore asting approa hes andtheir features . . . 39
2.3 Sevensubstation lients' ompositionsand orrelation oe ientswith
tem-peratures. . . 42
3.1 Fore asting result omparison between the naive model and thetimeseries
modelon the Substation VI_LOG data . . . 59
3.2 Fore asting result omparison between the naive model and thetimeseries
modelon the MVfeeder CLdata . . . 60
3.3 Performan e omparison among Time Series (TS) models with fore asting
temperature, a tual temperature, andnaive model. . . 64
4.1 9variables forthe average powerneural network model . . . 80
4.2 19variables for intradaypowervariation neural network model . . . 81
4.3 SubstationCE_MOU,fore astingresults: omparisonamongthenaivemodel,
time seriesmodel andNN models . . . 84
4.4 Substation CE_FRO, fore asting results: omparison between the naive
modeland theneuralnetworkmodel . . . 84
4.5 Summary of omparison aspe ts between neural network models and time
seriesmodels for the short-termload fore asting appli ation . . . 85
A.1 MV/LVsubstations,fore astingresults: omparisonbetweenthenaivemodel
andthe omplete TimeSeries (TS) modelof one-day-aheadfore asts. . . .155
A.2 MV/LVsubstations,fore astingresults: omparisonbetweenthenaivemodel
andthe omplete TimeSeries (TS) modelof two-day-aheadfore asts. . . .155
A.3 MV feeders, fore asting results: omparison between the naive model and
the ompleteTime Series(TS) modelof one-day-ahead fore asts.. . . .156
A.4 MV feeders, fore asting results: omparison between the naive model and
the ompleteTime Series(TS) modelof two-day-aheadfore asts. . . .156
D.1 6variables forthedailyaveragepowermodeland19 variables forthe
intra-daypowervariationmodel . . . .161
D.2 6variables forthe dailyaveragepowermodeland23 variables forthe
intra-daypowervariationmodel . . . .162
D.3 6variables forthe dailyaveragepowermodeland40 variables forthe
intra-daypowervariationmodel . . . .163
D.4 10 variables for the daily average power model and 37 variables for the
D.5 10 variables for the daily average power model and 37 variables for the
intraday powervariationmodel . . . .165
D.6 24 variables for the daily average power model and 32 variables for the
intraday powervariationmodel. . . .166
D.7 Substation CE_FRO:14variablesforthedailyaveragepowermodeland28
variables forthe intradaypower variation model. . . .168
E.1 Diérentshorizonsde temps pour laprévisionde harge . . . .174
E.2 Résumé desmodèles de harge prédi tifsetleurs ara téristiques . . . .177
E.3 Résumédela omparaisonentrelemodèleduréseaudeneuronesetlemodèle
A hieving myPhD degreehassetamilestone inmy life. The thesisdefensehasdrawnan
end to my best and worst moments during the three years that I have spent in G2elab.
People thathelpedmewill alwaysremaindear to mefor myfuturelife journey.
Above all, I would like to thank the reviewers and ommittee members of my thesis.
A very big thank to Pr. Carlo Alberto Nu i and Pr. Corinne Alonso for their timeand
energy to examine mydissertation and for their valuable opinions. Theiren ouragements
and appre iations give me strength and onden e inmyfuture work. Thanks also go to
Pr. DidierMayerforhis interests and insightful ommentsaboutmywork. Itwasagreat
honor for me to have Pr. Nouredine Hadjsaid as the president of the ommittee, and I
a knowledgehim for that.
I'malsogratefultotherepresentativesoftheindustrialpartnersofthe ompanythatI
workwith: Mr. OliverDevauxfromEDFandMr. AlainGlatignyfromS hneider Ele tri .
Ithankthemforthisinterestingsubje tthattheysetupandpertinent ommentsregarding
theindustrialappli ations ofmymodels.
I owe my sin ere gratitude to myprin ipal advisor, Pr. Yvon Bésanger. I would like
to thank him for his open-mind regarding ollaborations, for his support and trust, and
for hispatien e inguidan e. Ithank him for always being there for measa teammate at
di ulttimes thatwe en ountered throughpubli ations, and administrations.
IwouldliketoextendmygratitudetoDr. Frédéri Wurtzforhistrustandappre iation
for mywork,for hiswarmlywel oming meinto thelaboratory.
I would not forget to grant my gratitude to Pr. Gérard Dreyfus, Pr. Jean-Louis
La oume and Pr. Daniel Baudois for their s ienti guidan e. I highly respe t their
passion and rigorousness for the resear h. I thank them to a ept to oer me te hni al
advi eswithout reserve.
I want to expressmy gratitudeto Mr. ChristopheKeiny, Mr. Guillaume Antoineand
Miss Leti ia De-Alvaro from EDF for their energy devoting to my thesis proje t. They
have been supportive industrial advisers to keep me on tra k with the industrial needs,
and inthe meantimegive megreatfreedom to develop independent solutions.
I also would like to thank Mr. Frédéri Gorgette from ERDF, my supervisor of the
Mydeepest gratitude reserves to myfamily and my friends. Even though none ofmy
family member ouldattend mythesisdefense,their love and arearealways around me.
I want to thank myfriendsinG2elab fortheir toleran e andun onditional support tome
during those years. I have nevermet somany great people during so short period of my
life. I willtreasure our friendshipsfor alifelong time.
Youare responsible forwhat you have tamed. - The littleprin e
ADA Advan edDistribution Automation
ANN Arti ial Neuronal Network
AI Arti ialIntelligen e
AR AutoRegressive
ARMA AutoRegressive Moving Average
ARIMA AutoRegressive IntegratedMoving Average. Equation2.5
ARMAX AutoRegressive Moving AveragewitheXogenous inputs
ARIMAX AutoRegressive Integrated Moving Average witheXogenous inputs
ACF AutoCorrelationFun tion. Equation2.6
AFSA Arti ial FishSwarmAlgorithm
ANOVA ANalyse OfVArian e
ADF AugmentedDi key-Fullertest
AMR Automati MeterReading
CV Cross-Validation. Equation6.10
CDF Cumulated Distribution Fun tion. Equation3.13
CRLP Class Representative Load Pattern
CUSUM CUmulative SUM.Equation6.3
DSO Distribution SystemOperator
DMS Distribution Management System
DG Distributed Generators
DFT Dis reteFourier Transform. Equation3.6
DLE Distribution Load Estimation
ERDF Ele tri ité RéseauDistribution Fran e
EMS Energy Management System
FL FuzzyLogi
FCM Fuzzy C-Means
GDP Gross Domesti Produ t
GA Geneti Algorithm
GEV GeneralizedExtreme Value
GPD Generalized Pareto Distribution
HV/MV High Voltage/MediumVoltage
HV High Voltage
HW Holt-Winters
IA Immune Algorithm
ISODATA IterativeSelf-Organizing DATA-analysis te hnique algorithm
KNN K-Nearest Neighbor(s)
KPSS Kwiatkowski-Phillips-S hmidt-Shin tests
KDE KernelDensity Estimation. Equation 6.1
LV Low Voltage
LFC Load Frequen yControl
LTLF Long-Term Load Fore ast
LOO Leave-One-Out. Equation4.6
LL Lo alLinear. Equation 6.9
LL2 AdaptedLo alLinear
MV Medium Voltage
MV/LV Medium Voltage/ Low Voltage
MTLF Medium-Term Load Fore ast
MAPE Mean AbsolutePer entage Error. Equation2.32
MLE Maximum LikelihoodEstimation
MLP Multi LayerPer eptron
MAE Mean AbsoluteError. Equation2.33
MSE Mean Square Error
NN Neuronal Network
NARMA Nonlinear AutoRegressive Moving Average
NW Nadaraya-Watson. Equation 6.6
OLS Ordinary LeastSquare
PDF ProbabilityDensity Fun tion
PARMA Periodi AutoRegressiveMoving Average
PACF Partial AutoCorrelation Fun tion
PSO Parti le SwarmOptimization
PRESS Predi tedREsidualSumof Squares. Equation 4.6
PNN ProbabilityNeural Network
RNN Re urrent Neural Network
RBF Radial BasisFun tion
RBFN RadialBasis Fun tion Networks
RLP Representative Load Pattern
SE StateEstimator
STLF Short-Term Load Fore ast
SOM Self-Organizing Maps
SLP Single LayerPer eptron
SVM SupportVe tor Ma hine. Equation2.27
SVR SupportVe tor Regression. Equation2.27
SCADA Supervisory Control And DataA quisition
SSR SumofSquare Residuals
SSE SumSquare Error
TLP Typi al Load Prole
TMB Minimum Temperature Base
VVC VoltVAR Control
VLOO VirtualLeave-One-Out
VSTLF VeryShort-Term LoadFore ast
WCI Wind ChillIndex
WNN Wavelet Neuronal Network. Equation 2.18
WLSE Weighted LeastSquaresEstimation
x
Inuen evariableve toror variableto bedeterminedX
i
Sampled values, measurementsX
Observatonmatrix,whose elementx
ij
isthe measuredvalueof variablej
inexamplei
y
t
Load (power)measured at timet
y
Loadve tory
i
Sampled loadvalueǫ
t
Model'snoiseat timet
e
Dieren ebetween outputofthe modelandmeasured valueγ
y
(t, t − τ)
Auto ovarian e fun tion ofy
pro essat thetimet
andt
− τ
E
(⋅)
Expe ted value operator{P
i
, y
i
}
Histori al datainputs/outputspair, learningset or trainingset,i
= 1,⋯,N
f
t
Trendmodelvalueat timet
S
t
Cy li modelvalueat timet
D
α
, α
= 1,⋯,κ − 1
Dummyvariables,whereκ
isthenumberof dierent ategoriesγ
α
, α
= 1,⋯,κ − 1
Dummyregression oe ientsT
t
Temperature at timet
W
t
Detrendedseriesp
(ǫ)
Probability densityfun tion oftherandom variableǫ
F
ǫ
(x)
Cumulative distributionfun tion ofthe random variableǫ
P
Ve tor variables{p
j
, j
= 0,⋯,R}
of neural networks, whereR
is the total number of inputvariablesΩ
Setof theparameters of theneuralnetwork modelω
Ve tor ofweightsof the linear ombination, between thehidden layerand outputlayer of theneural networkmodelC
Ve tor{c
i
(P, Ω
i
), i = 1, ⋯, M}
oftheoutputsofhiddenneurons,whereM
isthenumber of hiddenneuronsr
0
Thethreshold rankof the orthogonalforward regressionr
probe
Therank ofa probevariableξ
i
Thei-th andidate variableve torf
(P
i
, Ω
)
Output of the neural network with respe t to the variable ve torP
i
and the parametersΩ
f
−i
(P
i
, Ω
)
Output of the neural network model when examplei
is withdrawn from the training setn
p
Numberof realizations ofthe probe variablen
rp
Number of realizations of the random probe whose rank is smaller than or equal to rankr
δ
Risk hosenbythedesignerto ontrol thenumberof inputsh
ii
Leverage, i-th diagonal element ofthe hat matrixH
p
Numberofsetof parametersof theneural networkmodel,whi hisequal to(R + 1)M +
(M + 1)
E
LOO
Leave-one-out s oreE
p
Approximationof the leave-one-out s oreZ
Ja obian matrix ofthe neural network modelE
yr
Yearlyenergy onsumptionE
0Non-heatingdailyenergy onsumption
s
Temperaturesensibility,indi ating theamountofenergy onsumed(kWh)byde reasing 1o
Ctemperature
E
n
Annualenergy onsumption adjusted tothe normal limati onditionP
(t)
Estimated meanpower attime tσ
(t)
Estimated standard deviation at hour tν
(t)
Estimated marginat timetb
(t)
Commongroup oe ientfortheBAGHEERAmodel onvertingheatingdailyenergy into heating powerat time tE
d
Daily energy onsumptionE
i
Meter re ordingenergy onsumption duringn
i
daysDd
i
Degree days duringa period ofn
i
daysDd
365Yearlydegree days inthe normal limati ondition
T
d
Dailyaverage temperatureT
N h
Nonheatingtemperature,atemperaturethresholdbelowwhi hthe onsumptionrises due tothe ele tri al heatersˆ
g
h
(x)
Kerneldensityestimator ofvariablex
,withsmoothing parameterh
h
SmoothingparameterK
(µ)
Normalkernel fun tionof variableµ
γ
Ex ess probability oftheload estimationmodelh
cv
Optimal smoothing parameterdened byCV te hniquey
T M B
_i
Estimated powerat TMB ondition
ˆ
f
h
′
(⋅)
Kernel-type estimator withitsoptimal smoothingparameterh
′
models in the smart grid ontext
Contents
1.1 Ba kground: smartgrid and smart meters for load modeling . 2
1.2 Motivation and obje tives . . . 3
1.2.a Fornetworkoperationneed . . . 4
1.2.b Fornetworkplanningneed . . . 5
1.3 Contributions of the thesis . . . 6
1.4 S ope and organization of the dissertation . . . 7
Abstra t
Ground-breakingevolutionshavebeenbroughttotraditionalele tri aldistributiongrids
by the on ept of smart grids. The smart meter system, as one of the most
impor-tant infrastru tures in the smart grids, gives us detailed information on ele tri ity
onsumptionofanindividual ustomer. In this ontext, weaimatdesigning
fore ast-ingmodelsandestimationmodels basedonthese informationfor needsindistribution
network operation andnetwork planning. The ontributions, aqui k overview of the
1.1 Ba kground: smart grid and smart meters for load
mod-eling
The smart-grid on ept ombines advan ed ommuni ation te hnologies with traditional
ele tri al distribution grids in order to improve the transparen y and the ontrollability
of distributiongrids. Fa ing several ground-breaking evolutions inthe ele tri ity systems,
su h as the large penetrations of the renewable power generation, the rapid load growth
due to plug-in ele tri vehi les, to name a ouple, numerous advan ed algorithms appear
in this ir umstan e to enhan e the stabilityand thee ien y of the system. These
Ad-van ed Distribution Automation (ADA) fun tions in lude Volt VAR Control (VVC) [1℄, self healing, and dire t load ontrol [2℄ (to name a few). The ADA fun tions are al u-lated in real-time or in ahead of time in order to help making de isions. Generally, the
monitoring and the ontrol pro ess ofdistribution networksareperformedat theMedium
Voltage (MV)level.
Oneofthesmart-gridgoalsisto makedistribution systemse onomi allye ientwith
reliableenergysuppliesandless osts. Distribution networkplanninginvolvesdevelopinga
s heduleoffutureadditionsthatensurethe qualityofenergydeliveryaswell asthelowest
possible ost. Ontheonehand,the ele tri ityinfrastru ture mustmeettheneedsof peak
loads. On the other hand, over-dimensioned systems an be very expensive. Thus,
reli-ableloadestimation models arerequiredto tightendistribution marginsand optimize the
planning investmentbyperformingdistributionnetwork al ulations, i.e., arryingout the
powerow al ulationin riti alsituations soasto identify poor ele tri itysupply zones.
Nevertheless,the omplexityinthe problemis relatedto theun ertaintyand randomness
inthe lients' ele tri ity onsumptions.
Inthe urrentstate,thes ar ityofmeasurementsonthedistributionsystemintrodu es
bottlene ks in arrying out theADA fun tions aswell asthenetwork optimization
al u-lations. Theavailable measurementsindistribution networksaremainly onthese ondary
of sour e substations. It is e onomi ally non-feasible to implement ele tri meters in all
738 000
Medium Voltage/ Low Voltage (MV/LV) substations. Today, for the operation need, applying very approximate probabilisti models with50% of pre ision seriouslyaf-fe ts the e ien y of the ADA fun tions, resulting dubious analysis results. In order to
omply with the planning need, the a tual model applied by the Fren h ele tri ity
om-pany, termedBAGHEERA, dependsmainly on the lient's individual information, whi h
be omes less and less available. Thus, a new model must be designed at the request of
repla ing theBAGHEERA model.
Starting from 2010, the Ele tri ité Réseau Distribution Fran e (ERDF) (Fren h
Dis-tribution System Operator (DSO)) laun hed the Linky (baptized name for the smart
meter inFran e) proje t, whi h aimsat installing
35 000 000
smart meters inFran e. On the one hand, end users will pay ele tri ity bills based on their real onsumptions ratherthan on the estimated ones as inthe today's ase inFran e. Onthe other hand, thanks
to these measurements, distribution network operators an have a better vision of the
substationsonthe Fren hdistribution networks. Intheexperimental phaseoftheLinky
proje t, the onsumption information of ea h individual is sampled on a 30-minute basis
andtransferredon eadaytothe orrespondentdata enter. However,asdataaregathered
inpa kagesandsentwitha ertainfrequen y[3℄,somedelayisfoundinthemeasurements.
Therefore,usingthea urateinformationprovidedbythesmartmeterstodevelopload
models isthe silverbulletthatmakeskeysmart-grid appli ationsfeasible.
1.2 Motivation and obje tives
The supervision of the power and voltage dispat hing of the networks is a riti al task
in distribution exploitation. It guarantees an e onomi al optimum and a dynami
sta-bility of the networks. Unlike transmission networks, on whi h abundant measurements
exist,distributionnetworkshavemu hlessmeasurements. Asamatteroffa t, be ause of
the omplex stru ture and a great number of nodes (MV/LV substations) indistribution
networks, itis e onomi ally impossible to install meters ina great quantity on these
sub-stationsindistributionnetworks. Thus,thedistribution systemis onsideredasblind or
nonobservable. Onesolutionto improve the observability of distributionnetworksis
to introdu e loadmodels inorder torepla e themeasurements.
In termsofloads indistribution networks, we distinguishtwo types:
MV lients dire tly onne ted toMV networks
Numerous Low Voltage (LV) lients onne ted to MV networks through the publi
MV/LV substations
HV
MV
LV
P,Q
|U
1
|,|I
1
|
|U
2
|,|I
2
|
|U
3
|,|I
3
|
|U
n
|,|I
n
|
...
Figure1.1: Availablemeasurements(markedinred)[4℄intheFren hdistributionnetworks.
∣ ⋅ ∣
representsthe norm notation,equivalent tothe magnitude.Currently,the measurements intheFren h distribution networksare(gure 1.1):
The a tive and the rea tive power on the se ondary High Voltage/Medium Voltage
(HV/MV)substations
The magnitudeof the urrent on head ofeveryMV feeder
The a tive andthe rea tive powerof some MV lients
OntheLV lients' side,the only available dataarelimited to thesubs ribed power in
the supply ontra t and billing information ofthe lients onne ted tothepubli MV/LV
substations.
With the new available individual onsumption data olle ted by smart meters, the
obje tive of the resear h program presented in this thesis is to build new load models
for the need in operation and in planning in distribution networks. This ontext makes
possible the design of a urate models for the distribution network planning, monitoring
and ontrol,inabsen e ofthe ostly measuring equipments indistribution networks.
1.2.a For network operation need
For the sake of ontrol and onguration in distribution systems, the evolution of the
MV/LV substation load needs to be known. Mainly, we an point out three dierent
reasonsdes ribedasfollows:
During a failure: in order to e iently restore ele tri ity in regions where a fault
o urs,loadsintheae tedregionsshould be knowninthefollowingthree minutes.
During network maintenan e: the variation of the onsumption needs to be known
to restorethepowersupply. Generally,atwo-dayperiodis onsidered bytheFren h
ele tri itydistributorERDFasanormalrepairing time. Inthis ase,atwo-dayload
fore astwithits standard error isneeded.
As inputs for the SE: the SE [5℄ is the ore fun tion of any energy management system. Itaimsat estimatingthenetwork variables,su hasthevoltage magnitudes
and angles. Figure 1.2 shows the s hemati of the relationship among fore asting
models,SE,andADAfun tions. Thefore asting modelsaswellasthenetworkdata
are onsidered as inputs for the SE. The network data [5℄ in ludes theinformation aboutthenetworktopology,lineresistan e,rea tan e, tapsetting, andline harging,
et . The output of the SE will lead the Distribution Management System (DMS)
ontrol s heduling blo k to perform on erned ADA fun tions for operational
de i-sions. These de isions enable the monitoring and ontrol of various devi es in the
networkssu has apa itor banks,Distributed Generators(DG), on-load tap
hang-ingtransformers,and swit hes/breakers, et .
The idea is then to design fore asting models for MV/LV substations relying on the
Linky
Linky
Linky
MV/LV Substation
Forecasting models
MV/LV Substation
Distributed Generation (DG)
Capacitor banks
HV/MV Substation
State estimator
Advanced Distribution Automation (ADA)
functions (VVC, self-healing...)
Optimized network reliability
Optimized network securities
Optimal economic benefit
Distribution Management System (DMS)
Figure1.2: Relationshipamongfore asting models, SE,andADA fun tions
1.2.b For network planning need
Designing a reliable distribution network is hallenging sin e it needs to guarantee a
sta-ble and ontinuous power supply to the ustomers. As a matter of fa t: a utility must
maintainthe voltage delivered toea h ustomer withina narrow range entered withinthe
voltages that the ele tri equipment is designed totolerate [6℄. In theEuropean ele tri ity regulation, for the LV networks, a
10%
out-of-range voltage is a eptable. Beyond this range, the ustomeris dened asapoorlysupplied ustomer.For the sake of planning, network al ulations are performed under extreme
situa-tions in order to handle worst ase s enarios [7, 8℄. More spe i ally, these al ulations are arriedout when either of these two ases o urs: maximum demand withminimum
generation, and maximum generation with minimum demand [9℄. With the arrival of a onsiderable portion of DGs into the networks, the later ase an be expe ted. Be ause
ofthe ustomers' various behaviors,ina samegeographi zone, peak demandsof dierent
loadpattern ateveryhour isrequired. Un ertaintyoftheloadestimationalsoneedstobe
onsidered [10℄. Generally, for the voltage-drop al ulation, the ex ess probability, whi h denes the threshold power bounds, is xed to 10% [11℄. Consequently, in the network planning, a lient's total supply demand in ludes the lient's daily load pattern and his
10%ex ess probabilityun ertainty.
Moreover,inthe distributionnetwork planningpro ess,alistofstandardsand riteria
for equipments, su h asdistribution lines and MV/LV transformers,to namea ouple, is
alsodened. Reliableloadmodelsarealsorequiredinthefollowing ases inorderto arry
out distribution network al ulations:
Distribution line losses
Currents inlinesegments
Networkvoltage-drops
MV/LVtransformers: two-hourequivalentpower 1
,voltage-drops,andele tri allosses
Therefore,the obje tive isto dene an individual ustomer's maximumand minimum
loadlimits ofa yearwith10% ex essprobability.
1.3 Contributions of the thesis
Themajorpurposesofthisdissertationaretwofold: designingnewloadfore astingmodels
for thenetwork operation and load estimation models for thenetwork planning in
distri-bution networks. The ontributions of thethesis an be summarizedasfollows:
Load fore ast is an extensively investigated subje t on the transmission level [12,
13, 14, 15,16,17, 18,17,19, 20℄. However on the distribution level, withthe same hara teristi onsumption data (several dozen kW), to the best of our knowledge,
few workshave been done.
We an think of three plausible reasons explaining this fa t: rst, more attention
has been paid to the transmission level, asthe transmission grid extents on longer
distan es and overs larger territories. The transmission grid, involving huge osts,
is theba kbone of the powersystems. Se ond, thehighervoltage level onsumption
has a more regular load urve pattern, whi h makes it easier to fore ast. Third,
there were no available measurements on the MV/LV substations relying on whi h
thefore asting models an be designed and validated.
In this resear h proje t, two models based on time series and neural networks have
been proposed,for thenetwork operation need.
The dierent load types (residential, ommer ial, and industrial) are examined in
this resear h,and their dierent properties arepointedout.
1
The two methods are designed and evaluated on real measurements olle ted from
theFren hdistributiongridsintheframeworkoftheLinky proje t. Referen e ase
isestablishedinorderto be ompared to the twoproposed models. Advantagesand
drawba ksofthe twomethods aredrawnthroughthe omparison. Thetwomethods
arealternative and,at thesame time, ompliment to ea hother to setthepre ision
limit dueto theintrinsi hara teristi s ofthesubstation loaddata.
Timeseriesmethodpresentedinthisdissertationisanoriginalwork,wherenumerous
statisti altoolsareintegratedinforthepursuing ofpre ision. Residualofthemodel
islooked through indetailsto ensure the model'swell being.
Neural network method, on the other hand, is inspired by the sele tion pro edure
proposedbyGérardDreyfus[21℄,arenownedexpertintheneuralnetworkmodeling. Wefo usontheNeuronalNetwork(NN)modeldesign,whi hhasbeenfullyexploited
for the rsttime inthe short-term loadfore ast.
Until theimplementationofthesmartmeters, usuallytherewerenoavailable
histor-i alloaddatabesidesthedemand surveydataonalimitednumberof lients. Inthe
loadmodeling eld,most of theworks on erningthedistributionnetwork planning
aimatestimatingthepeakdemandforagroupof ustomersduringthepeakdemand
of the system, namely the oin ident peak demand [6℄. For this purpose, some re-sear hesbasedonend-usemethod[9,22℄de omposetheloadmodeloftheresidential ustomerintoapplian eelementaryunits. Othersfo usonthe lassi ationmethod,
sorting ustomersinto dierent ategories, and on the representation ea h ategory
witha Typi al Load Prole (TLP). In the smart metering ontext, we are therst
to propose the on ept of individual data-driven load model for ustomers, for the
network planning need.
Tobuildanindividualestimationmodel,therelationshipbetweentheele tri ity
on-sumptionandthetemperatureisdedu edbynonparametri estimators. Themethod
isappliedtoreal onsumptiondataofindividual ustomersinFran e. Performan eis
omparedtothe urrentloadmodel(termedBAGHEERA)oftheele tri ity ompany
EDFthrough dierent validationstudies.
1.4 S ope and organization of the dissertation
Thedissertation is organizedasfollows:
Chapter1statesthatthe smartgrid andthe smartmetersareinrapid development
to improve the e ien y andthe ontrollabilityof distribution networks. The
revo-lutionary hangesindistributionnetwork systemsurgethea urate loadfore asting
models. Re ordingindividual onsumptioninformation,thesmartmetersenablethe
buildingofthesemodels. Thethesisisen ouragedbytheLinkyproje tlaun hedby
theFren hele tri itydistributorERDF,aimingatinstalling
35 000 000
smartmeters inFran e. Twomainobje tivesareevokedinthesmartmetering ontext: short-termnetwork planning. Reasonsfor thedevelopment ofthese two obje tivesarede lared.
Contributions ofthe thesis arehighlighted.
The rest of the dissertation is divided into two parts, dealing with two distin t
ob-je tives. Part A ta kles the fore asting load models for the operation need, and in ludes
hapters 2,3and 4.
Chapter 2 gives a review of the load fore asting methods in literatures and set up
basi framework for the performan e evaluation. Load fore asts are stratied into
dierentobje tivesa ordingtotheirleadtimes ale. Dierentobje tivesaredevoted
to dierent appli ations. More input information is needed for a longer lead time
fore ast. Fore astingmethodsaredividedinto two ategories: the lassi alapproa h
and the arti ial intelligent approa h. Hybrid models that possess the advantages
of both ategories gain more and more popularity in the appli ations. Data used
inour study arethoroughly analyzed, suggestinginuen e fa torsto our short-term
load fore asting models. We argue for our hoi es of timeseries and neuralnetwork
methods. Performan e riteria anda referen emodel(termednaive model) arealso
established in this hapter, building a solidframework base for the presentation of
theapplied methods inthe nexttwo hapters.
Chapter 3 presents the short-term load fore asting model based on the time series
method. The fore asting pro edure is detailed. The additive time series model
ontains three omponents: a trend, a y li and a random error. The rst two
omponents aredeterministi , and are designed into models respe tively. Thetrend
modelistemperature-dependent, linear,anddummyvariablesintegrated,indi ating
day types. Cy li model is omposed by the Fourier omponents, whose
frequen- ies are found by a smoothed periodogram. Numerous statisti al tools are applied
pursuing a better pre ision: sliding windowstrategy isadopted sothatthemodelis
updatedduringea hfore astingperiod;ANalyseOfVArian e(ANOVA)nullitytest
is applied to estimatethesigni an e ofvariables. Availabledataaredivided into a
learning set and a test set. Important parameters of the model are dened thanks
to thelearning set. Whereasthe test set isreservedfor the performan e evaluation.
Residual is examined, making sure the well t of the models. Weather un ertainty
impa ton the pre isionof thefore asting modelisdis ussed intheendof the
hap-ter. We on luded thatevenwith theweather un ertainty,our proposed timeseries
modelstill outperforms thenaive model.
Chapter 4 introdu es the neural network model from the arti ial intelligent
ap-proa h family. In our study, we fo us on the design of the neural network model,
hoosingtheoptimalmodelthathasthebesta hievablepredi tiveability. Ageneral
on ept of the ma hine learning te hnique is stated, and di ulties su h asnding
therelevant variablesand thebias-varian e dilemma,areexplained. Theorthogonal
forward regression and the Virtual Leave-One-Out (VLOO) te hnique areproposed
as solutions to these di ulties. Experiments on the same data show that the
pro-posed methodology behaves better than the time series model in term of a ura y.
A omparison indiverspropertiesismadebetween thetwomodelsintheend ofthis
Part B introdu es the load estimation problem for the planning need, and proposes
solutions. It ontains hapters 5 and 6.
Chapter 5 fo uses on the load resear h proje ts in distribution networks. Load
re-sear h proje ts aimat providinghourlyloadestimationmodels for individual lient.
Three steps, i.e., te hni al analysis, e onomi al analysis, and nal de ision, for the
de ision makings in distribution network planning are presented. The outputs of
the load resear h proje ts ontribute to the te hni al analysis, devoting to nding
solutions in network planning. The me hanism of the aggregation of loads, the
o-in ident load, is explained. Common methods ofnding TLP,whi h represents the
dailyloadpattern of a ertaingroup of lients, arepresented. Load models usedby
ele tri al unities in Finland, Denmark, Norway, and Taiwan aredes ribed. Finally,
the omponents ofthe BAGHEERA modelapplied bytheFren h DSOare detailed
andthe method isdemonstrated withreal measurements.
Chapter6proposesanovelapproa hfortheindividualloadestimationinthe ontext
of smart meters. With the abundant individual onsumption information, in our
opinion,the loadmodelis readyto beindividualized rather thanestimated through
the TLPs. Thus, inthis hapter, theindividual loadestimation modelbasedon the
nonparametri estimators isput forward. Numerous statisti altools,su h asbinary
hypothesis tests, kernel densityestimation, CUmulative SUM (CUSUM) algorithm,
andCross-Validation(CV)te hnique,areintegrated intheproposedmethod. Three
kernel regressors, i.e., Nadaraya-Watson (NW), Lo al Linear (LL), and Adapted
Lo al Linear (LL2) are applied to dedu e the relationship between the load and
the temperature variations. Dierent appli ation ases a ording to thequalityand
quantityofthedataaresuggested. Themethodisillustratedwithrealmeasurements
and ompared withtheBAGHEERAmodel. Thevalidityofthemethodisexamined
withextensive examples. In the end ofthe hapter, adis ussion onthedenition of
the un ertaintyboundof theestimationmodelis arriedout.
Chapter7 on ludesthedissertationandproposesperspe tivesforthefutureresear h
Short-term load fore asting models
framework
Contents
2.1 Literature review . . . 14
2.1.a Fore astingleadtimesandinuen e fa tors . . . 14
2.1.b Fore astingmethods . . . 16
2.1.b-i Classi alapproa h. . . 18
2.1.b-ii Arti ialintelligentapproa h . . . 25
2.1.b-iii Hybridmodels . . . 35
2.1. Literaturereview on lusionsandperspe tives . . . 37
2.2 Data des ription. . . 40
2.2.a MV/LVsubstation . . . 40
2.2.a-i Temperatureinuen e . . . 41
2.2.a-ii Daytypeinuen e . . . 42
2.2.a-iii Timeinuen e . . . 42
2.2.b MVfeeder. . . 45
2.3 Choi es of Time series and NN methods. . . 45
2.4 Performan e riteria and referen e ase. . . 46
2.4.a Performan e riteria: MAPEandMAE . . . 46
2.4.b Referen e ase: thenaivemodel. . . 47
2.5 Con lusion . . . 47
Abstra t
Load fore ast plays an important role in de ision makings in power systems. This
hapter begins with the review of load fore asting models in literatures. The se ond
partof the hapter ontributes toaframework onsisting ofdata des ription,method
sele tion,performan e riteria,andreferen e aseintrodu tions. Inthereviewingpart,
we lassify a wide range of approa hes of load fore ast into two ategories: lassi al
approa h,andarti ialintelligentapproa h. Methodologiesinea h ategoryarebriey
presented. Theiradvantages,disadvantages,appli ations andpertinentresear hworks
arealsodeveloped. Popularhybridmodels ombiningtwoormoredierentapproa hes
arealso involved. In the framework part, datausedfor the design andthe evaluation
of our methodologies are analyzed. Certain behavioral omponents in the data are
pointedout. The hoi es ofthe methodologies basedon the timeseriesandtheneural
networksareargued. Thesetwomethodologies aregoingtobedetailedinthe following
two hapters. Performan e riteria and referen e asearestated soas tolay agood
2.1 Literature review
The quality of the de ision making in ele tri power systems strongly depends on the
a ura y of the power load predi tions. Various de isions require reliable and a urate
loadfore asting models withdierent time-s alesaswell asondierent hierar hi al levels
innetwork systems [23℄.
A wide range of approa hes have been proposed to the load fore asting problems. In
thisse tion,weaimatpresentingbrieythedierentapproa hesfoundinliteratures,their
spe i ities,appli ations and te hniquesapplied to loadfore ast.
Theorganization of the se tion isasfollows: rst,we start byintrodu ingdenitions,
appli ations and inuen e fa tors of dierent lead time load fore asts. Then, a two
di-mensional digestinlead timeand involtage hierar hy s alessummarizes load fore asting
methods,followedbythe des riptive presentation ofeverymethod. Related worksarealso
depi ted. Finally,we on lude theliterature reviewinatable.
Notethatfor the sake of larityandease ofunderstanding, mathemati al notations in
the referen ed works and internal reports have been adapted in order to keep oheren e
through theentiredissertation.
2.1.a Fore asting lead times and inuen e fa tors
Dierent fore asting lead times result into dierent fore asting models as well as their
inputvariables. Numerousfa tors,su hasweather onditions,seasonalee ts,andso ial,
e onomi ,demographi fa torsexplainthevariationsintheload[23℄. Table2.1summarizes theappli ationsandinuen e fa torsfor dierent timehorizon fore asting models[23,24,
18℄.
Noti ethat more input variables arein luded when thetime horizon be omeslonger.
ForaVSTLF,univariate(onlythehistori alpowersamplesare onsideredasinputs)
mod-els an oersatisfa toryresults. TheseVSTLFsoftenparti ipateto improvethee ien y
and reliabilityofthe real-timeele tri al systems. For longerlead timefore ast,
multivari-atemodelswithexogenousvariablesarefavored. TheSTLF needsmainlythree ategories
of inputs: weather, alendar, and histori al variables [25℄. Dueto some measurement de-lays, or the omputational timefor the exe ution of theADAfun tions, the STLFs often
repla e VSTLFs to fulllneedsin network operations. STLF fore asts also help redu ing
equipment failures and system bla kouts by indi ating the operational margins in power
systems. TheMTLFmodelshelpmakingnan ialde isions,su hasevaluationofthepri e
of energy produ ts and investment interests. In su h ases, the fore asting models need
additional inputs asso ial and e onomi al fa tors. The LTLF, on erning energy system
apitalexpendituresandmoreimportante onomi investments, needstotakeintoa ount
more so io-e onomi fa tors,and sometimes even their futureevolutions.
The herein des ription is in a general way, as in se tion 2.2.a, we will talk about an
industrialMV/LVsubstationloadthatisindependentto weather onditions. Thus,
what-everthefore asting leadtime is, for this loadexample,weather variableis not onsidered
asan inuen e fa tor.
Asdes ribed in the table 2.1, for our appli ation in network operations, espe ially to
ooperatewiththeADAfun tions,wefo usontheSTLF. Weather, alendarandhistori al
Table 2.1: Dierent timehorizon loadfore asts
Timehorizon Appli ations Inuen efa tors
Very Short-Term Load
Fore ast (VSTLF)
(1Min
∼
1h)ADA fun tionsinDMS,Load
Frequen y Control (LFC) in
Energy Management System
(EMS)
Histori al onsumptions
Short-Term Load
Fore- ast (STLF) (1h
∼
1 week)Operation (ADA fun tions),
estimation of load ows,
rep-resentation ofsavingpotential
for e onomi and se ure
oper-ationof powersystems
Histori al onsumptions,
al-endar fa tors (day type and
houroftheday),weather
on-ditions (
∗
)Medium-Term Load
Fore ast (MTLF) (1
week
∼
1 year)Negotiation of ele tri ity
on-tra ts, s heduling of fuel
sup-plies and maintenan e
opera-tion
(
∗
) + population, e onomi fa tors, et (◇
)Long-Term Load
Fore- ast (LTLF) (1 year
∼
severalyears)Capital expenditures and
planning operations
(
◇
) + more information su h as: population growth, GrossDomesti Produ t(GDP)
(
∗
) and (◇
) represent respe tively inuen e fa torsfor short-term and medium-term fore- asts.For a weather sensitive load, in ludingthe redible fore asting weather information as
input is re ommended as it an improve the performan e of the fore asting model. T
em-perature and humidity are the most frequently used load predi tors. Composite weather
variables su h asTemperature-Humidity Index (THI),Wind Chill Index(WCI) [24℄, and smoothed weather variables [26℄ are often adopted. THI and WCI indi ate respe tively thedis omfort ausedbysummerheatandwinterwind hill. Thesmoothedweather
vari-able represents the ee ts of hanges in weather a umulated over the time. In pra ti e,
regardingSTLF,weather fore asting data areapplied to al ulate theperforman eof the
model [27℄. However, for the most of the time, in the onstru tion phase of the model, thepredi tedweatherfore astisnot available. Inthis ase, most authorsintheload
fore- astingeldrunsimulationswiththe realizedweather data[28℄. Oneshould bearinmind thatusingfore astingweatherinformationwillsurelyde reasethemodel'soverallpre ision
[26,29℄. Therefore,someauthors[30℄proposedomittingimpre ise weatherinformation as a onservativesolution, sin e itwouldbringlargevarian eto themodel. Apromising way
to handle the un ertainty in weather variables is the weather ensemble predi tions that
generatetheloadfore astsinaprobabilisti form[29℄. In hapter3,wewillre-dis ussand showempiri al eviden eaddressing to thisissue.
The alendar inputs in lude the time of the year, the day of the week, the hour of
the day as well as day types (working days, weekends or national holidays). There are
oftenmoredi ulttofore astthanworkingdaysduetotheirrelativeinfrequento urren e
and the lients' irregular behaviors. Someauthors workon the lassi ation methods [19℄ inorder to ndor even reate similar days [31℄for these anomalous dayfore asts.
Histori aldatainputsarealsoveryimportant totheSTLF,fromwhi htheseasonality
information an beextra ted [14℄.
2.1.b Fore asting methods
This subse tiongives anoverviewof various approa hes for loadfore asts. Manyof them
aredevelopedforSTLF onthe HighVoltage(HV)level,althoughMVandLVlevelsbegan
to attra t more attention with the expansion of the smart grids during the past years.
Figure 2.1 gives a two dimensional digest on the methods and the models in the load
fore asting eldboth intheleadtimeand thevoltage hierar hi al s ales.
Mainly,two lasses ofapproa hes anbedistinguished [14,23℄: lassi alapproa h and Arti ialIntelligen e(AI) approa h. Classi alapproa hrequiresanexpli itmathemati al
modelwhi hinterpretstherelationshipbetween loadanditsinuen efa tors. Thisfamily
in ludes regressionmodel, timeseries method, similardayapproa h, end-use method and
e onometri approa h. AI approa h, on the other hand, extra ting non linear
relation-ships between input fa tors and load hasbe ome very popular nowadays. This family of
algorithms in ludesArti ial Neuronal Network (ANN),fuzzylogi , and expertsystems.
Very short-term
Short-term
Medium-term Long-term
H
V
M
V
L
V
ANN
(12)
End use method
(13)
ANN (14)
(MAPE 4.9-7.74%
depending on the
forecast season)
ANN (11)
(Inputs: 1. Historical load data
2. Estimated load pattern)
C
lassification method
(Client typical pattern)
+ANN (15)
SVM
(8)
Similar days+NN correction model
(9)
Expert system
(10)
FL or Fuzzy Neural Networks
(FNN) (7)
ANN
(6)
Regression method
(4)
Multivariable time series
ARIMAX (3)
Univariable time series:
(1) Exponential smoothing
(2)Box-Jenkins (AR, ARMA,
ARIMA)
Econometric approach
(5)
Figure 2.1: Summary of load fore asting methods in two dimensions: time horizon and
voltagehierar hy. HV:HighVoltage,MV:MediumVoltageandLV:LowVoltage. Numbers
appearinthe gure orrespond tothe relatedworks. a
a
(1):[13,14℄(2):[13,15,14℄(3):[16℄(4):[17,18,17℄(5):[24℄(6):[19,20℄(7):[32,33,34℄(8):[35,36℄(9):[31℄ (10):[37,38℄(11):[39℄(12):[22℄(13):[40℄(14):[41℄
b
Univariatetimeseriesmodelreferstothemodelwithonlyoneobservationseries,i.e.,loaddata.
c
2.1.b-i Classi al approa h
Regression model. Regression is one of the most widely used statisti al te hniques.
Ele tri load fore asting regression methods are usually used to express the relationship
between load onsumption and external fa tors[24,42℄:
y
i
= a
i
x
i
+ e
i
(2.1)where
y
i
is the i-th load sample,x
i
is the inuen e variable ve tor orrespondent to the i-th load sample,a
i
is the transposed regression oe ient ve tor, ande
i
is a Gaussian error.The advantages of regression methods are relatively easy implementation and
inter-pretation for the relationship between input and outputvariables. Another advantage of
the method is that it is easy to ompute thepredi tion interval through estimation error
of the model. The disadvantage of regression methods is the need to identify a orre t
form in luding ee tive inputs and output. This is hard due to the omplex non-linear
relationship [23℄.
C.L.Horetal. [43℄developedseveralmultipleregressionmodelsin orporating weather-relatedandso io-e onomi variablesontheloaddemandforEnglandandWales. Monthly
data from 1989 to 1995 are usedfor the oe ient estimation of the model and monthly
data from 1996 to 2003 are used to evaluate the a ura y of thefore asting model. The
non-linearrelationshipbetweenweather-relatedfa tors(meanmonthlytemperaturevalue)
and load suggests introdu ing other weather omposite variables, su h asHeating Degree
Days (HDD),Cooling Degree Days(CDD),and EnthalpyLatent Days(ELD).The
so io-e onomi variable GDP has evidently an impa t on the trend. One of their regression
modelis set up as:
E
ˆ
1
= ( ˆ
E
A
+ α
7
GDP)F
adj
(y)
, whereE
ˆ
1
is the predi ted ele tri ity de-mand,ˆ
E
A
= α
0
+ α
1
CDD+ α
2
HDD+ α
3
ELD+ α
4
V
w
+ α
5
M
s
+ α
6
M
r
,whi h represents the weather-related model.V
w
stands for the mean monthly wind speed,M
s
stands for the mean monthly sunshine hours, andM
r
stands for themonthly rainfall.α
n
, n
= 0,⋯,7
are onstant oe ients.F
adj
(y)
is the adjustment fa tor for ea h year. The Mean Abso-lute Per entage Error (MAPE) of themodel, whi h represents theaverage portion of theabsolute fore asting errorsto therealfore asting values,wasaround 2%.
A. Bruhns et al. [17℄designed a non-linear regression model for MTLF. This hourly loadpredi tionmodeltermedEventail hasbeenappliedbytheFren hele tri ity ompany
EDF sin e 2001. They de omposite the load
P
i
into three omponents:P
i
= Phc
i
+
P c
i
+ e
i
, whereP c
i
andP hc
i
are respe tively the weather-dependent and the weather-independent parts,e
i
isa Gaussian error. Theweather-dependent partis ttedbya non-linear model of the observed temperature, the exponential smoothing temperature, andthe loud over. Two thresholds for heating and ooling temperatures arealso adopted in
orderto opewiththenon-linearity. Theexponentialsmoothingontemperatureree tsthe
inertiaoftemperatureinsidebuildingsto theoutsidetemperature variation. The
weather-independentpartintegratestrends,day,week,yearperiods,anddaytypeinformation. The
dummyvariables are usedto indi ate daytypes and four termsof Fourier series areused
to model the seasonality pattern. Despite of the omputational di ulty in estimation
dueto thetemperaturesmoothingparameters,thresholds,andstrongnon-linearities,they
de lared that with the known weather data, a MAPE of around 2% for one-year-ahead
W. Charytoniuk et al. [26℄ introdu ed a nonparametri regression model for STLF. They onsidered load as a weighted average of past loads. The spe i weights are
de-ned by amultivariate kernel and its smoothing parameters. The optimal values ofthese
smoothing parameters are al ulated using CV te hnique. A onditional expe tation of
theload is built based on thelo al neighborhood loads
P
i
:P
ˆ
(x) =
∑
n
i
=1
{P
i
∏
r
j
=1
K
(
xj −xji
hj
)}
∑
n
i
=1
{∏
r
j
=1
K
(
xj −xji
hj
)}
,where
h
1
,
⋯, h
r
aresmoothingparametersandK
(µ)
isanormalkernel.j
denotesthej-th inuen e variable in the exogenous variable ve torx
andi
denotes the i-th observation sample. They have obtained with a urate temperature values, up to one-week-ahead, aMAPE of2.78% ompared to an ANNmodel of2.64%. Theyarguedthateven theerrors
areslightly higherthan theANN,the errorsobtained bytheregressionmodelhave better
properties ompared to those oftheANN model.
Time series method. Timeseries methodisalinearmodelbasedontheassumption
that there existsan internal stru ture within data. This stru ture onsists of
auto orre-lation, trend and periodi variations. Time series methods dete t and explore these
rela-tionshipsbetweenthe urrent loadvalue,histori aldataandsometimes exogenousfa tors.
A distin tive property oftime series models is to onsidertime asone of theexplanatory
variables. As a linear model, time series method is suitable for VSTLF and STLF and
it is able to provide the predi tion interval. Whereas for MTLF and LTLF, these
inter-nalrelationships are no longer linear and themodel be omes ompli ated. Thus, itis no
longer e ient to use a linear timeseries model for the MTLF or the LTLF appli ation.
Inparti ular, Box-Jenkinsmodels andexponential smoothing modelsarethemost widely
usedtime series methods.
1. Box-Jenkins models [44℄in lude AutoRegressive (AR), Moving Average (MA), Au-toRegressive Moving Average (ARMA), AutoRegressive Integrated Moving A
ver-age (ARIMA), Periodi AutoRegressive Moving Average (PARMA), and
AutoRe-gressive Moving Average with eXogenous inputs (ARMAX). They are adapted for
STLF, sin e they allow for theexpli it modeling of timedependen e. The basi
el-ement of these models are AR and MA models. AR model aims at estimating the
urrent loadbythemoving averageofhistori alload. MAmodel,ontheotherhand,
explains the urrent loadwiththe pasterrors ommittedbythemodel.
Insome ases,wherethestationary ondition 1
isnotmet,thetrendandtheperiodi
ee ts an be removed by arrying out transformationsasfollows:
Let
y
t
be asto hasti pro ess. Witharst dieren e,thelineartrend isremoved:Z
t
= y
t
− y
t
−1
(2.2)Withase ond dieren e,quadrati trendis removed:
Z
t
= (y
t
− y
t
−1
) − (y
t
−1
− y
t
−2
)
(2.3)Forperiodi series,itispossibletoadjustobservationseasonalee t by arryingout
1