ENERGY CENTER State Utility Forecasting Group (SUFG) Methods for Forecasting Supply and Demand...

ENERGY CENTERState Utility Forecasting Group (SUFG)


Methods for Forecasting Supply and Demand

Presented by:Douglas J. GothamPurdue University

Presented to:Institute of Public Utilities

56th Annual Regulatory Studies Program

August 12, 2014



2

Using the Past to Predict the Future

• What is the next number in the following sequences?0, 2, 4, 6, 8, 10, ….

0, 1, 4, 9, 16, 25, 36, ....

0, 1, 2, 3, 5, 7, 11, 13, ....

1, 3, 7, 15, 31, ....

0, 1, 1, 2, 3, 5, 8, 13, ....

8, 6, 7, 5, 3, 0, ….

8, 5, 4, 9, 1, 7, ….



3

A Simple Example

900

920

940

960

980

1000

1020

1040

1060

1080

1100

1 2 3 4 5 6

1000

1010

1020

1030

1040

1050

?

?

?



4

A Little More Difficult1000

1100

1210

1331

1464

1610

?

?

?

900

1000

1100

1200

1300

1400

1500

1600

1700

1 2 3 4 5 6



Much More Difficult18831

18794

18193

19944

20855

20858

19275

19054

20315

21002

?

?

17,000

18,000

19,000

20,000

21,000

22,000



6

Much More Difficult• The numbers on the previous slide were

the summer peak demands for Indiana from 2002 to 2011

• They are affected by a number of factors– Weather– Economic activity– Price– Interruptible customers called upon– Price of competing fuels



7

Question• How do we find

a pattern in these peak demand numbers to predict the future?

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

2010

0

5000

10000

15000

20000

25000



8

Methods of Forecasting

• Palm reading• Tea leaves• Tarot cards• Ouija board• Crystal ball• Polling

• Astrology• Dart board• Sheep entrails• Hire a consultant• Wishful thinking



9

Alternative Methods of Forecasting

• Top-down– trend analysis (aka time series)– econometric

• Bottom-up– survey-based– end-use

• Hybrid– statistically-adjusted end-use



10

Time Series Forecasting• Linear Trend

– Fit the best straight line to the historical data and assume that the future will follow that line

• works perfectly in the 1st example

– Many methods exist for finding the best fitting line; the most common is the least squares method

Y X



11

Time Series Forecasting• Polynomial Trend

– Fit the polynomial curve to the historical data and assume that the future will follow that line

– Can be done to any order of polynomial (square, cube, etc.) but higher orders are usually needlessly complex

21 2 ...Y X X



12

Time Series Forecasting• Logarithmic Trend

– Fit an exponential curve to the historical data and assume that the future will follow that line

• works perfectly for the 2nd example

XY



13

Example

• Use linear time series analyses to project Indiana peak demand from 2010 to 2029 using historical observations over 3 time periods– 1980-2009– 1990-2009– 2000-2009



14

Trend 1 - Starting in 1980

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

2016

2018

2020

2022

2024

2026

2028

2030

0

5000

10000

15000

20000

25000

30000

Actual

Trend



15


1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

2016

2018

2020

2022

2024

2026

2028

2030

0

5000

10000

15000

20000

25000

30000

Actual

Trend



16


2000

2002

2004

2006

2008

2010

2012

2014

2016

2018

2020

2022

2024

2026

2028

2030

0

5000

10000

15000

20000

25000

30000

Actual

Trend



17

Comparison

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

2010

2012

2014

2016

2018

2020

2022

2024

2026

2028

2030

10000

15000

20000

25000

30000

Actual

Trend 1

Trend 2

Trend 3



18

ResultsYear Trend 1 Trend 2 Trend 3

2010 20981 20963 20793

2015 22766 22716 22376

2020 24551 24470 23959

2025 26337 26224 25542

2030 28122 27978 27124



19

Advantages• Relatively easy• The statistical functions in most

commercial spreadsheet software packages will calculate many of these for you

• Requires little data



20

Disadvantages

• Does not account for changing circumstances

• Choice of historical observations can impact results

• May not work well when there is a lot of variability in the historical data– If the time series curve does not perfectly

fit the historical data, there is model error



21

Acceptability

• Trend analysis was a popular forecasting methodology until the 1970s

• The inability to handle changing conditions led to considerably inaccurate forecasts

• They have been largely discredited– MISO’s forecasting whitepaper lists it as an

“unacceptable” method



22

Econometric Forecasting• Econometric models attempt to quantify the

relationship between the parameter of interest (output variable) and a number of factors that affect the output variable.

• Example– Output variable– Explanatory variable

• Economic activity• Weather (HDD/CDD)• Electricity price• Natural gas price• Fuel oil price



23

Estimating Relationships• Each explanatory variable affects the output

variable in a different way. The relationships (or sensitivities) can be calculated via any of the methods used in time series forecasting– Can be linear, polynomial, logarithmic, moving

averages, …

• Relationships are determined simultaneously to find overall best fit

1 1 2 2 3 3 ...Y X X X



24

A Simple Example• Suppose we have 4

sets of observations with 2 possible explanatory variables

Output Y Variable X1 Variable X2

110 100 100

113 120 110

114 130 90

121 150 120

80 90 100 110 120 130 140 150 160100

110

120

130

80 90 100 110 120 130 140 150 160100

110

120

130



25

A Simple Example

• Including both variables provides a perfect fit– Perfect fits are not usually achievable in

complex systems

Y = 0.2X1 – 0.1X2 + 100



26

Advantages• Improved accuracy over trend analysis• Ability to analyze different scenarios• Greater understanding of the factors

affecting forecast uncertainty



27

Disadvantages

• More time and resource intensive than trend analysis

• Difficult to account for factors that will change the future relationship between the drivers and the output variable– utility DSM programs– government codes and standards



28

Acceptability

• Econometric methods became popular as trend analysis died out in the 70s and 80s

• They continue to be used today• MISO’s forecasting whitepaper lists it as

an “acceptable” method



29

Survey-Based Forecasting

• Also referred to as “informed opinion” forecasts

• Use information from a select group of customers regarding their future production and expansion plans as the basis for a forecast

• Commonly done with large users



30

Advantages

• Simplicity• The ability to account for expected

fundamental changes in customer demand for large users, especially in the near-term– new major user or customer closing a

facility



31

Disadvantages

• Tend to be inaccurate beyond first few years– most customers do not know what their

production levels will be five or ten years in the future

– few customers expect to close shop– new customers after first couple years are

unknown• Lack of transparency



32

Acceptability

• Survey-based forecasts may be acceptable for short-term applications or if used in conjunction with another method in the longer term

• MISO’s forecasting whitepaper lists it as an “unacceptable” method



33

End Use Forecasting• End use forecasting looks at individual

devices, aka end uses (e.g., refrigerators)• How many refrigerators are out there?• How much electricity does a refrigerator use?• How will the number of refrigerators change

in the future?• How will the amount of use per refrigerator

change in the future?• Repeat for other end uses



34Source: Van Buskirk, Robert. “History and Scope of USA Mandatory Appliance Efficiency Standards.” (CLASP/LBNL).



35

Advantages• Account for changes in efficiency levels (new

refrigerators tend to be more efficient than older ones) both for new uses and for replacement of old equipment

• Allow for impact of competing fuels (natural gas vs. electricity for heating) or for competing technologies (electric resistance heating vs. heat pump)

• Incorporate and evaluate the impact of demand-side management/conservation programs



36

Disadvantages

• Tremendously data intensive• Primarily limited to forecasting energy

usage, unlike other forecasting methods– Most long-term planning electricity

forecasting models forecast energy and then derive peak demand from the energy forecast



37

Acceptability

• End-use modeling was first developed in the 1970s but started to gain popularity with the increase in DSM in the 1990s

• MISO’s forecasting whitepaper lists it as an “acceptable” method



38

Hybrid Forecasting

• Hybrid models employ facets of both top-down and bottom-up models

• Most common is called the statistically-adjusted end-use (SAE) model

• In reality, most end-use models are hybrid to some degree in that they rely on top-down approaches to determine the growth in new devices



39

SAE Models

• SAE models incorporate features of both econometric and end-use models

• Adjust the end-use estimated loads using a statistical regression to match observed loads



40

Disadvantages

• Increased model complexity• More time and resource intensive



41

Advantages

• In general, hybrid approaches attempt to combine the relative advantages and disadvantages of both model types

• Can better capture externalities that affect customer decisions when compared to end-use models– green options



42

Acceptability

• Hybrid models have been gaining in popularity in recent years

• MISO’s forecasting whitepaper lists it as an “acceptable” method



43

Forecasting Example

• SUFG has electrical energy models for each of 8 utilities in Indiana

• Utility energy forecasts are built up from sectoral forecasting models– residential (end-use & econometric)– commercial (end-use & econometric)– industrial (econometric)



44

Another Example

• SUFG is developing independent forecasting models for MISO– econometric– individual state level (15 states)



45

Another Example• The Energy Information Administration’s

National Energy Modeling System (NEMS) projects energy and fuel prices for 9 census regions

• Energy demand (end-use)– residential– commercial– industrial– transportation



46

Sources of Uncertainty• Exogenous assumptions

– forecast is driven by a number of assumptions (e.g., economic activity) about the future

• Stochastic model error– it is usually impossible to perfectly estimate the

relationship between all possible factors and the output

• Non-stochastic model error– bad input data (measurement/estimation error)



47

Energy → Peak Demand• Constant load factor / load shape

– Peak demand and energy grow at same rate

• Constant load factor / load shape for each sector– Calculate sectoral contribution to peak demand

and sum– If low load factor (residential) grows fastest, peak

demand grows faster than energy– If high load factor (industrial) grows fastest, peak

demand grows slower than energy



48

Energy → Peak Demand

• Day types– Break overall load shapes into typical day

types• low, medium, high• weekday, weekend, peak day

– Adjust day type for load management and conservation programs

– Can be done on a total system level or a sectoral level



Load Diversity• Each utility does not see its peak demand at

the same time as the others• 2011 peak demands occurred at:

– Duke Energy – 7/20, 2 PM– Hoosier Energy – 7/21, 6 PM– Indiana Michigan – 7/21, 2 PM– Indiana Municipal Power Agency – 7/21, 2 PM– Indianapolis Power & Light – 7/20, 3 PM– NIPSCO – 7/21, 4 PM– SIGECO – 7/21, 4 PM– Wabash Valley – 7/20, 8 PM

• Statewide peak – 7/21, 4 PM



50

Load Diversity Example

1AM

2AM

3AM

4AM

5AM

6AM

7AM

8AM

9AM

10AM

11AM

12PM

1PM

2PM

3PM

4PM

5PM

6PM

7PM

8PM

9PM

10PM

11PM

12AM

0

500

1000

1500

2000

2500

20002000 2000

A

B

C



51

Example (continued)

1AM

2AM

3AM

4AM

5AM

6AM

7AM

8AM

9AM

10AM

11AM

12PM

1PM

2PM

3PM

4PM

5PM

6PM

7PM

8PM

9PM

10PM

11PM

12AM

0

1000

2000

3000

4000

5000

60005700

Total



52

Load Diversity• This analysis is normally performed for

all hours of the year• Thus, the statewide (or regional) peak

demand is less than the sum of the individual peaks

• Actual statewide/regional peak demand can be calculated by summing up the load levels of all utilities for each hour of the year



53

Diversity Factor

• The diversity factor is an indication of the level of load diversity

• Historically, Indiana’s diversity factor has been about 96 – 97 percent– that is, statewide peak demand is usually

about 96 percent of the sum of the individual utility peak demands



54

Resource Expansion Models• Determine amount, type, and location of

demand/supply resources to be developed to reliably meet future electricity demand

• Detailed information on existing generation resources but limited detail on existing transmission resources

• EGEAS, NEEM, NEMS, NESSIE, ReEDS, Strategist



55

Production Costing Models

• Simulates the network operation over a year to determine costs, emissions, impacts of congestion

• Detailed information on generation and transmission infrastructure (both existing and future)

• GE MAPS, GRIDVIEW, PROMOD, UPLAN



56

Power Flow Models• Determine adequacy of transmission system

to meet demand without violating physical constraints (undervoltage, overcurrent, etc.)

• Examine contingencies (can system operate reliably if a line/transformer/generator goes down?)

• Load/generator output fixed for a given hour• Models laws of physics that drive actual path

of power flow (not an optimization)• Power World, PSS/e, PSLF



57

Further Information

• State Utility Forecasting Group– http://www.purdue.edu/dp/energy/SUFG/

• Energy Information Administration– http://www.eia.doe.gov/index.html

http://www.purdue.edu/dp/energy/SUFG/

http://www.eia.doe.gov/index.html

ENERGY CENTER State Utility Forecasting Group (SUFG) Methods for Forecasting Supply and Demand...

Documents

Transcript of ENERGY CENTER State Utility Forecasting Group (SUFG) Methods for Forecasting Supply and Demand...