Data Mining techniques application in Power Distribution ...ewh.ieee.org/conf/tdc/PRESENTATION_PES08_08TD0649.pdf · IEEE PES TD 2008 CLASSIFICATION Rule set for the working days

Data Mining techniques application in Power Distribution Utilities

Authors:Sérgio RamosZita A. Vale

GECAD – Knowledge Engineering and Decision Support R esearch GroupEngineering Institute of Porto – Polytechnic Institu te of Porto

IEEE PES TD 2008Chicago, IL, USA22th April 2008

IEEE PES TD 2008

INTRODUCTION

Content of Presentation:

� Electricity Market Liberalization Environment

� MV Costumers Characterization – An overview

� Data Mining Techniques Application

� Clustering and Consumers Characterization• Case Study

� Future Work

2

IEEE PES TD 2008

ELECTRICITY MARKET LIBERALIZATION

� Total freedom in choosing the electricity supplier

� Consumers and suppliers are exposed to price risk

� Distribution and retail companies are looking for better tariff rates

� Competitive environment among retail companies to sell the electricity

3

IEEE PES TD 2008


� Increase of demand elasticity due to the electricity price volatility

� Electricity customers more concerned about their consumption behaviour

� Knowledge about customers’ daily load profile is essential for leadership in this new context

� Deeper relationship between Customer and Electricity Supplier

4

ElectricitySupplier

Consumer

IEEE PES TD 2008


CONSUMERS’ CHARACTERIZATION� Advantages:

�Design of new tariffs, contracts, products and services

�Creation of incentive actions to the energy efficiency

5

IEEE PES TD 2008

MV COSTUMERS CHARACTERIZATION

DETERMINATION AND CHARACTERIZATION OF MV CONSUMERS LOAD PROFILE USING DATA MINING

TECHNIQUES

- Rule Sets- Decision Tree- Overall Accuracy

Relationship consumer/Electricity Supplier

Data BaseData Mining Techniques

New TariffSchedules

ClustersData Pre-

processing

Formatted Data

ClassificationModel

- C5.0 Class. Algor.- Shape indices

ClusteringAlgorithms

- Two-Step- K-Means- SOM

Typical LoadProfile

6

IEEE PES TD 2008

LOAD STUDY

Data description:

� Sample of 229 MV Consumers

� Collect period of data(3 months in the Summer / Winter – from the Portuguese Distribution Company)

� Consumed power recorded with a cadence of 15 minutes

� 96 values obtained per day

ll (m(m)) = {l= {l 11(m)(m), , …… , l, l9696

(m)(m)} } with m = number of customerswith m = number of customers

7

IEEE PES TD 2008

DATA PREPARATION

Data-Cleaning

� 21 customer’s files were discarded� Some damaged files were detected� Customers without registered values� 208 customers remained to be analyzed

� To estimate missing values of measures a multi laye r perceptron – MLP – artificial neural net was used

� The errors of the metered load curves are attenuate d without making significant alterations in the real measures

8

IEEE PES TD 2008

DATA PREPARATION

Pre – processing data:

� The Power consumption was normalized to the [0-1] range

� A representative load diagram has been built for each customer by averaging the related load diagrams

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

9

TREATED

DATA

WEEK

DAY - YEAR

WEEKEND

DAY - YEAR

REPRESENTATIVE DIAGRAMS:

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Time (h)

Pow

er(p

.u.)

IEEE PES TD 2008

CLUSTERING PROCESS

REPRESENTATIVELOAD DIAGRAMS


CLUSTERSCLUSTERS

NUMBEROF

CLUSTERS

NUMBEROF

CLUSTERS

- TWO-STEP- K-MEANS

- SOM

ClusteringPerformanceComparison

10

Choiceof the bestAlgorithm

IEEE PES TD 2008

CLUSTERING PROCESS

Mean Index Adequacy (MIA):

Cluster Dispersion Index (CDI):

∑

∑ ∑

=

= =

=K

k

k

K

k

n

n

kmk

RrdK

LldnK

CDI

k

1

)(2

1 1

)()(2)(

),(2

1

),(.2

11)(

),(1 )(

1

)(2 kK

k

k LrdK

MIA ∑=

×=

Choice of the Clustering Algorithm:

11

IEEE PES TD 2008

CLUSTERING PROCESS

Comparison of the clustering performance:Two-Step Cluster AnalysisK-meansKohonen Net – Self Organizing Features Maps

Two-StepK-means

SOM

MIA

CDI0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

MIA

CDI

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

3 6 9 12 15 clusters

MIA

CDI

12

IEEE PES TD 2008

TWO-STEP CLUSTERING APLICATION� Using the Two-step cluster algorithm the clusters were

obtained using the representative load diagrams

13

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Working days Time (h)

Pow

er (

p.u.

)

IEEE PES TD 2008

REPRESENTATIVE LOAD DIAGRAMS

� Representative diagram for each cluster

14

Weekend days

0,0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Time (h)

Pow

er (p.

u.)

Work days

0,0

0,10,2

0,30,4

0,5

0,60,7

0,80,9

1,0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Time (h)

Pow

er (p

.u.)

IEEE PES TD 2008

CLASSIFICATION MODEL

Objective:

� To build a classification model, that applied to ne w unclassified records, will allow to foresee the cla ss to which it belongs

� In the future it will allow to attribute to each ne w consumer the consumption profile that best represents it

15

IEEE PES TD 2008


C5.0 Algorithm

� Decision Tree:• Application simplicity• Result in tree form• Generation of rules

Derive from the daily load diagrams� Give information about:

• The daily load curve shape• The consumption pattern of each consumer

dayav

day

P

Pf

,

min,3 =

day

dayav

P

Pf

max,

,1 =day

day

P

Pf

max,

min,2 =

dayav

nightav

P

Pf

,

,

3

14 =

dayav

lunchav

P

Pf

,

,

8

15 =

16

IEEE PES TD 2008


Classification module framework:

17


- C5.0 Classification Algorithm


GENERATION OF RULESDECISION TREE

LOAD SHAPE INDEXES(Each representative load curve is represented by a set of load shape

indexes)

[ ]654321 ,,,,, fffffff =

- INPUT ATTRIBUTES: VECTOR {f}- TEST SET- TRAINING SET- TEN-FOLD CROSS VALIDATION

- EVALUATION ACCURACY

- ANALYSIS OF THE CONFUSION MATRIX

IEEE PES TD 2008

CLASSIFICATION

� Rule set for the working days classification model:

� Overall accuracy:

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 5 ≤ 0,55 and f 1 ≤ 0,35 and f 4 ≤ 0,31 then cluster 8

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 5 ≤ 0,55 and f 1 ≤ 0,35 and f 4 > 0,31 then cluster 9

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 5 ≤ 0,55 and f 1 > 0,35 then cluster 5

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 5 > 0,55 and f 2 ≤ 0,06 then cluster 7

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 5 > 0,55 and f 2 > 0,06 then cluster 6

if f 3 ≤ 0,48 and f 2 > 0,13 and f 4 ≤ 0,24 then cluster 4

if f 3 ≤ 0,48 and f 2 ≤ 0,13 and f 4 > 0,24 then cluster 5

if f 3 > 0,48 and f 3 ≤ 0,78and f 2 ≤ 0,44 then cluster 3

if f 3 > 0,48 and f 3 ≤ 0,78and f 2 > 0,44 then cluster 2

if f 3 > 0,48 and f 3 > 0,78 then cluster 1

94,83%

18

IEEE PES TD 2008

CONSUMER-SUPPLIER RELATIONSHIP

� The Knowledge can be used by the Retail Companies

• Identify diagrams’ peaks

• Develop specific consumer's contracts

• Optimization of the offers of electric power purchase

19

IEEE PES TD 2008

CONSUMER-SUPPLIER RELATIONSHIP

� The Knowledge can be used by the electric power consumers

• Choice of the electricity supplier with the best tariff schedule proposal

• Modulation of their electric consumption habits

• In the execution of electric energy interruption contracts

20

IEEE PES TD 2008

FURTHER WORK

� Compare the efficiency of the C5.0 algorithm with different classification algorithms

� The design of new prices categories in order to adequately adapt the tariff schedules to the cluster consumption pattern

� Formulation of new tariffs schedules in articulation with electricity markets

21

Data Mining techniques application in Power Distribution Utilities

Authors:Sérgio RamosZita A. Vale

GECAD – Knowledge Engineering and Decision Support R esearch GroupEngineering Institute of Porto – Polytechnic Institu te of Porto

IEEE PES TD 2008Chicago, IL, USA22th April 2008

Documents

Data Mining techniques application in Power Distribution ...ewh.ieee.org/conf/tdc/PRESENTATION_PES08_08TD0649.pdf · IEEE PES TD 2008 CLASSIFICATION Rule set for the working days