IEEE PES General Meeting 2019 | 4 -8 Aug 2019, Atlanta, GA ... · Low-level aggregation -> Active and reactive power are mutually informative at low-level aggre-gation, so a multivariate

1

IEEE PES General Meeting 2019 | 4-8 Aug 2019, Atlanta, GA USAPanel Session on Probabilistic Energy Forecasting

PROBABILISTIC INDUSTRIAL LOAD FORECASTING

Guido CarpinelliDepartment of Electrical Engineering and

Information TechnologyUniversity of Napoli Federico II – Naples, Italy

Antonio Bracale, Pasquale De FalcoDepartment of Engineering

University of Napoli Parthenope – Naples, Italy

2

Contents1. Introduction

2. Methodologies

3. Numerical applications

4. Conclusions

3

IntroductionIndustrial load is a big share of the total electrical load

36.4%

36.8%

41.5% 46.3%

26.4%43.8%

39.6%

33.8% 38.5%

33.5%

39.4%

30.2%

EU28 EU19 Austria Belgium France Germany Italy NL Poland Spain Sweden UK

0

500

1000

1500

2000

2500

3000

Elec

trica

l ene

rgy

cons

umpt

ion

[TW

h]

Total energy consumption

Industrial energy consumption

Ref. Eurostat Data for 2016

4

IntroductionMuch of the industrial load is non-controllable and characterized by an intrinsicrandomness due to human behavior, non-schedulable activities, and contingencies

Forecasts of industrial load allow:TSO and DSO for dispatching energy and for acquiring energy reserves

5

IntroductionMuch of the industrial load is non-controllable and characterized by an intrinsicrandomness due to human behavior, non-schedulable activities, and contingencies

Forecasts of industrial load allow:Ownership of the industrial system for bidding on markets and for managing DERsand equipment in industrial smart grids and microgrids

6

IntroductionTraditional (regional) load forecastingSeasonality, calendar events, and weather are common features in modeling andforecasting load

Industrial load forecastingTraditional load forecasting methods may not adapt well to industrial load, due to thedifferent nature and peculiar characteristics of industrial sites

Major challenges: individuation of work regimes, low dependency on ambienttemperature, and low-level load aggregation!!

7

IntroductionChallenge 1: Individuation of work regimesSome industrial load follow regime-switching profile patterns due to turn-on andturn-off of large machines. This determines heteroscedasticity, i.e., conditionalvariance cannot be assumed constant throughout the entire sample

Forecasting models based on assumptions upon low variability of conditionalresiduals might be inappropriate to fit industrial load

Berk K, Hoffmann A, Muller A. “Probabilistic forecasting of industrial electricity load with regime switching behavior,”Int. J. Forec., vol. 34, 2018

8

IntroductionChallenge 2: Low dependency on ambient temperature

-20 -10 0 10 20 30 40

Ambient temperature [°C]

1000

2000

3000

4000

5000

6000

7000

8000

Load

zon

e 8

[kW

]

-20 -10 0 10 20 30 40

Ambient temperature [°C]

0

20000

40000

60000

80000

100000

120000

Load

zon

e 9

[kW

]

GEFCom2012 data

Traditional load Industrial load

9

IntroductionChallenge 3: Low-level load aggregationCompared to load at regional or national level of aggregation, industrial load followpatterns that are much less smooth. At individual industrial-load level, patterns mightalso be intermittent, due to the need for manual usage by operators

2016-09-19 2016-09-21 2016-09-23 2016-09-25

Hour of the week [h]

0

1

2

3

4

5

6

7

8

9

10

Indi

vidu

al in

dust

rial l

oad

[kW

]

2016-09-19 2016-09-21 2016-09-23 2016-09-25


0

40

80

120

160

200

240

280

320

Aggr

egat

e in

dust

rial l

oad

[kW

]

2016-09-19 2016-09-21 2016-09-23 2016-09-25


0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Aggr

egat

e re

gion

al lo

ad [M

W]

ISO NE data Data of an Italian industrial site

10

Methodologies

Industrial point (16.2%)

Industrial probabilistic (3.0%)

Commercial probabilistic (2.0%)

Commercial point (25.3%)

Residential probabilistic (5.1%)

Residential point (48.5%)

Despite allocating the most of the share of the total aggregate load, industrial load forecasting is not as popular as residential or commercial load forecasting

Papers published since 2010

19.2%

27.3%

53.6%

11

MethodologiesRelevant Probabilistic Industrial Load Forecasting (PILF) methodologies:

• Sparse Heteroscedastic Gaussian Process• Regime-switching Markov Chains with time series• Quantile Regression Forest with electric predictors• Multivariate Quantile Regression on individual forecasts

Example of an application of PILF:• Management of a distribution transformer by dynamic transformer rating

12

MethodologiesQuantile Regression Forest (QRF)Challenge 1:Work regimes -> QRF models allow clustering similar observations in specific leaves

Bracale A, Carpinelli G, De Falco P. “Comparing univariate and multivariate methods for probabilistic industrial load forecasting,” in Proc. of EFEA2018, Rome, 24-26 September 2018

13

MethodologiesQuantile Regression Forest (QRF)Challenge 1:Work regimes -> QRF models allow clustering similar observations in specific leaves

Leaf containing past observations

Path determined by split on predictors


14

MethodologiesQuantile Regression Forest (QRF) with electric predictorsChallenge 2:Low dependency on ambient temperature -> Exploit “new” variables that are informative for the

target variable!

Active and reactive power are mutually informative in industrial frameworksOther electric variables (e.g., voltage), together with calendar variables and information onmanufacturing schedules and work shifts within the industrial site might be informative forpredicting the load (and even forindividuating work regimes!!!)

Bracale A, Caramia P, De Falco P, Hong T. “Short-term industrial reactive power forecasting,” Int. J. Electr. Power Energy Systems, vol. 107, 2019

15

MethodologiesQuantile Regression Forest (QRF) with electric predictorsActive and reactive power are individually forecasted by QRF in a univariate approach,exploiting informative electric predictors


16

MethodologiesMultivariate Quantile Regression (MQR) on individual forecastsChallenge 3:Low-level aggregation -> Active and reactive power are mutually informative at low-level aggre-

gation, so a multivariate scheme may increase the skill of forecast

Bracale A, Caramia P, De Falco P, Hong T. “A multivariate approach to probabilistic industrial load forecasting,” submitted to Int. J. Electr. Power Energy Systems

17

MethodologiesMultivariate Quantile Regression (MQR) on individual forecastsChallenge 3:Low-level aggregation -> Active and reactive power are mutually informative at low-level aggre-

gation, so a multivariate scheme may increase the skill of forecast

Models can be re-trained as new observations become available, in order to catch dynamicvariation of industrial load pattern on the basis of recent outcomes

Multivariate approaches are computationally intensive, but MQR models are easily re-estimated as new observations become available, since the underlying optimization problem(Pinball Score minimization during the training stage) can be set in a linear programming form

Bracale A, Caramia P, De Falco P, Hong T. “A multivariate approach to probabilistic industrial load forecasting,” submitted to Int. J. Electr. Power Energy Systems

18

MethodologiesApplication of PILF: dynamic transformer ratingIn an industrial microgrid, forecasts of the industrial load may be used to manageDERs and to operate lines and transformers at their dynamic rating

Dynamic Transformer Rating (DTR) allows exploiting the interface transformer at its best, by avoiding the Hottest-Spot Temperature (HST) going beyond dangerous levels

19

MethodologiesApplication of PILF: dynamic transformer rating

Bracale A, Carpinelli G, De Falco P. “Probabilistic risk-based management of distribution transformers by dynamic transformer rating,”Int. J. Electr. Power Energy Systems, vol. 113, 2019

20

Numerical applicationsExperiments are carried out on an Italian industrial site dedicated to powertransformer manufacturing

MV

LV

Aggregate loadSingle feederPainting machine

21

Numerical applications1-hour forecasts99 predictive quantiles are provided for each target time horizonPinball Score (PS) and Coverage Error (CE) are shown to assess the performance

-15.5% -18% -20.5% -17.5% -29.4% -23.5%

22


1 24 48 72 96 120 144 168

Hour [h]

0

50

100

150

200

250

300

350

400

450

Load

[kW

]

98% prediction interval




actual load

23


0 0.2 0.4 0.6 0.8 1

Nominal coverage [-]

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Estim

ated

cov

erag

e [-]

Forecast

Ideal

24

Numerical applicationsApplication: dynamic transformer rating

MODEL

AGGREGATE LOADActive power Reactive powerPS

[kW]CE [%]

PS [kVAr]

CE [%]

QRF 522.76 2.77 371.85 2.12

Forecasts of the aggregate industrial load can be used to operate the feeding transformer at its dynamic rating

Active power forecasts

Reactive power forecasts

Transformer load forecasts

25


Winter monthFebruary 2018

Load and ambient temperature forecasts

obtained by QRFs

1 48 96 144 192 240 288 336 384 432 480 528 576 624 672

Hour of the month [h]

100

200

300

400

500

600

700

800

900

1000

1100

Tran

sfor

mer

cur

rent

[A]

Unmanaged transformer load

Low acceptable risk

High acceptable risk

26


𝑘𝑘𝑟𝑟 = 1Positive

outcomesNegative outcomes

Positive predictions 182 44Negative predictions 29 417𝐴𝐴𝐴𝐴𝐴𝐴 [-] 0.891𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡

(𝑚𝑚) [€] 389.33𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡

(𝑖𝑖𝑖𝑖) [€] 110.11𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡

(𝑢𝑢) [€] 985.59𝐸𝐸𝐸𝐸𝑡𝑡𝑡𝑡𝑡𝑡

(𝑚𝑚) [MWh] 3.44𝐸𝐸𝐸𝐸𝑡𝑡𝑡𝑡𝑡𝑡

(𝑖𝑖𝑖𝑖) [MWh] 8.80

𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡(𝑚𝑚): economic impact with QRF forecasts

𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡(𝑖𝑖𝑖𝑖): economic impact with ideal forecasts

𝐴𝐴𝑡𝑡𝑡𝑡𝑡𝑡(𝑢𝑢): economic impact of unmanaged transformer𝐸𝐸𝐸𝐸𝑡𝑡𝑡𝑡𝑡𝑡

(𝑚𝑚): energy not delivered with QRF forecasts𝐸𝐸𝐸𝐸𝑡𝑡𝑡𝑡𝑡𝑡

(𝑖𝑖𝑖𝑖): energy not delivered with ideal forecasts

0 2 4 6 8 10 12 14 16 18 20

Acceptable risk level

0

100

200

300

400

500

600

700

800

900

1000

Econ

omic

al im

pact

of t

he lo

ss o

f life

[€]

Managed transformer

Unmanaged transformer

0 2 4 6 8 10 12 14 16 18 20

Acceptable risk level

0

1.5

3

4.5

6

7.5

9

10.5

12

13.5

15

Mon

thly

ene

rgy

unde

liver

ed [M

Wh]

Managed transformer

Unmanaged transformer

27

Conclusions• Even if most of the electrical energy is consumed by industries, industrial load

forecasting is often overlooked, specially within probabilistic frameworks

• In order to reach excellence, PILF methodologies must account for individuation of work regimes, for low dependency on ambient temperature, and for low-level aggregation

• Univariate and multivariate methods have been applied with success to an Italian industrial load, exploiting information provided by electrical and industrial-process-related predictors. Improvements are in the range 15-30%

• Applications of PILF to DTR show potentialities in managing components at their maximum capability

28

THANK YOU FOR YOUR ATTENTION!!!

Slide Number 1Slide Number 2Slide Number 3Slide Number 4Slide Number 5Slide Number 6Slide Number 7Slide Number 8Slide Number 9Slide Number 10Slide Number 11Slide Number 12Slide Number 13Slide Number 14Slide Number 15Slide Number 16Slide Number 17Slide Number 18Slide Number 19Slide Number 20Slide Number 21Slide Number 22Slide Number 23Slide Number 24Slide Number 25Slide Number 26Slide Number 27Slide Number 28

Documents

IEEE PES General Meeting 2019 | 4 -8 Aug 2019, Atlanta, GA ... · Low-level aggregation -> Active and reactive power are mutually informative at low-level aggre-gation, so a multivariate