OVERVIEW OF LOAD FORECASTING METHODOLOGY Northeast Utilities Economic & Load Forecasting Dept. May 1, 2008 UConn/NU Operations Management.

OVERVIEW OF LOAD FORECASTING

METHODOLOGYNortheast UtilitiesNortheast Utilities

Economic & Load Forecasting Dept.Economic & Load Forecasting Dept.

May 1, 2008May 1, 2008UConn/NU Operations ManagementUConn/NU Operations Management

Forecast PurposeForecast Purpose

Provide an Provide an indicationindication of expected sales volumes 1 to of expected sales volumes 1 to 5 years out, given certain assumptions5 years out, given certain assumptions Considered “most likely” with equal chance of too high or too Considered “most likely” with equal chance of too high or too

lowlow Therefore, Company should plan for a range of possible Therefore, Company should plan for a range of possible

outcomesoutcomes

General guidance regarding sales trends, based on General guidance regarding sales trends, based on economic theory and available dataeconomic theory and available data

Used primarily for financial forecasting & rate casesUsed primarily for financial forecasting & rate cases Some use for transmission and supply planningSome use for transmission and supply planning

Forecast TheoryForecast Theory

Economic theory drives the forecasting structure and Economic theory drives the forecasting structure and modelsmodels

Consumption is a function of a primary economic Consumption is a function of a primary economic driver, price of the product, price of competing driver, price of the product, price of competing products, and a vector of other relevant variablesproducts, and a vector of other relevant variables

Theoretical structure is critical to withstand scrutiny of Theoretical structure is critical to withstand scrutiny of forecast reviewforecast review Particularly important in rate case processParticularly important in rate case process Becoming more important within corporationBecoming more important within corporation

Forecast PracticeForecast Practice

Good forecast must have good inputsGood forecast must have good inputs Accurate billed vs. calendar sales, and timely billing and Accurate billed vs. calendar sales, and timely billing and

booking of salesbooking of sales Accurate customer countsAccurate customer counts Accurate and relevant economic dataAccurate and relevant economic data

Forecasts are dynamic and will vary with each Forecasts are dynamic and will vary with each forecast solutionforecast solution Updated historical data (internal and external)Updated historical data (internal and external) Relationship of sales to their drivers (elasticities) updatedRelationship of sales to their drivers (elasticities) updated New forecasts for economic, price, C&LM, ED, customer New forecasts for economic, price, C&LM, ED, customer

specific, etc.specific, etc. Model differences and changes in customer behaviorModel differences and changes in customer behavior

Forecast MethodologiesForecast Methodologies

Combine the strengths of multiple methodologies:Combine the strengths of multiple methodologies:

Primarily “End-Use” Models Primarily “End-Use” Models Sales = #Customers * Use per CustomerSales = #Customers * Use per Customer

Within various customer classes and end-useWithin various customer classes and end-use Enables capturing of structural changes in demandEnables capturing of structural changes in demand However, extremely data intensiveHowever, extremely data intensive

Blended with Econometric Models (“causal linear regression”)Blended with Econometric Models (“causal linear regression”) y = a + bxy = a + bx1 1 + cx+ cx22 + … + …

Far less data intensive (time series of y and x’s)Far less data intensive (time series of y and x’s) But, assumes historic relationships will continueBut, assumes historic relationships will continue

Supplemented by JudgmentSupplemented by Judgment AE input on customer specific changesAE input on customer specific changes C&LM and ED impactsC&LM and ED impacts Analyst judgmentAnalyst judgment

Industrial (net of Special Contracts)

Where to Begin?Where to Begin?

Annualized Industrial Gas SalesAnnualized Industrial Gas Sales

PSNH Normalized Residential Calendar Sales 12 Month Ending

2600

2700

2800

2900

3000

3100

3200

3300

Jan-02 Jul-02 Jan-03 Jul-03 Jan-04 Jul-04 Jan-05 Jul-05 Jan-06 Jul-06 Jan-07 Jul-07 Jan-08 Jul-08

GWH

Normalized Actual 2008 Budget

We’d Like to Begin Here!We’d Like to Begin Here!

END-USE MODELSales = # of Units * Use per Unit

GENERIC MODEL EXAMPLE - RESIDENTIAL

Forecast Total Market Residential Customers = 1,000,000

Forecast End-Use Saturations - % of Total Market with End Use

25% of Residential Customers Have Electric Water Heaters

Forecast Usage per End Use

Electric Water Heater Usage = 4,000 KWH/Year

Sales = Total Market * Saturation * Usage

Sales = 1,000,000 * 25% * 4000 KWH = 1000 GWH

Customer ForecastCustomer Forecast

y = #Customersy = #CustomersXX11 = Housing Starts = Housing Starts

XX22 = Own Price = Own Price

XX33 = Competing Price = Competing Price

XX44 = lag(#Customers) = lag(#Customers)

y = a + bxy = a + bx11 + cx + cx22 + dx + dx33 + ex + ex44

Need both historic and forecasted time series for y and each xNeed both historic and forecasted time series for y and each x Regression run over historic time period (say 1990 – 2007)Regression run over historic time period (say 1990 – 2007) Solve equation over forecast time period (say 2008 - 2013)Solve equation over forecast time period (say 2008 - 2013)

Graphing the Data Always Helps!!!Graphing the Data Always Helps!!!

Residential Customers and Connecticut Permits

0.0%

1.0%

2.0%

3.0%

Perc

en

t C

han

ge in

Resid

en

tial C

usto

mers

0

5,000

10,000

15,000

20,000

25,000

30,000

CT

Perm

its

Residential Customers CT Permits

End-Use Saturation ForecastEnd-Use Saturation Forecast

Energy Information Administration (EIA) Energy Information Administration (EIA) historic regional datahistoric regional data

Adjusted based on Company-specific Adjusted based on Company-specific Customer SurveysCustomer Surveys

Trend model applied to create forecasted Trend model applied to create forecasted saturationssaturations

Usage ForecastUsage Forecast

Usage is adjusted by a Price Elasticity estimate:Usage is adjusted by a Price Elasticity estimate:

y = Use per Customery = Use per CustomerXX11 = Own Price = Own Price

XX22 = Income = Income

XX33 = Competing Price = Competing Price

XX44 = lag(#Use per Customer) = lag(#Use per Customer)

y = a + bxy = a + bx11 + cx + cx22 + dx + dx33 + ex + ex44

estimate of “b” is used to develop price elasticityestimate of “b” is used to develop price elasticity Base Usage is adjusted throughout the forecast based Base Usage is adjusted throughout the forecast based on forecast of change in price times elasticity estimateon forecast of change in price times elasticity estimate

End-Use EquationEnd-Use Equation

Sales =Sales =(Customers * Saturation EU1) * (Base (Customers * Saturation EU1) * (Base Usage EU1 * Price Elasticity Impact) + Usage EU1 * Price Elasticity Impact) +

(Customers * Saturation EU2) * (Base Usage (Customers * Saturation EU2) * (Base Usage EU2 * Price Elasticity Impact) + ……..EU2 * Price Elasticity Impact) + ……..

Across multiple end-uses for each of Residential, Commercial and Across multiple end-uses for each of Residential, Commercial and IndustrialIndustrial

Economic DriversEconomic Drivers

• EmploymentEmployment• Personal IncomePersonal Income• Housing Starts or StockHousing Starts or Stock• Gross State ProductGross State Product• Manufacturing Worker HoursManufacturing Worker Hours• Industrial ProductionIndustrial Production• InflationInflation

Other Forecast InputsOther Forecast Inputs

• Appliance Efficiency Standards Appliance Efficiency Standards • Company-sponsored Programs (DSM, Company-sponsored Programs (DSM,

Economic Development)Economic Development)• Weather (forecast assumes “normal” Weather (forecast assumes “normal”

weather)weather)• Vulnerable LoadVulnerable Load• Self-generationSelf-generation• Large Customer SurveysLarge Customer Surveys

How to Assess Forecast ResultsHow to Assess Forecast Results

• Degree of confidence in quality of data inputsDegree of confidence in quality of data inputs• Degree of confidence in model diagnostics (e.g., Degree of confidence in model diagnostics (e.g.,

regression stats)regression stats)• Changes in annualized sales to show trends in Changes in annualized sales to show trends in

growthgrowth• Assess YTD growth in salesAssess YTD growth in sales• Look for patterns in Large Customer’s usageLook for patterns in Large Customer’s usage• Residential customers and Use per Customer trendsResidential customers and Use per Customer trends• Economic AssessmentEconomic Assessment• Monitor industry trends, technology trends, efficiency Monitor industry trends, technology trends, efficiency

trendstrends• Look at ISO-NE Load Forecast Comm. survey results Look at ISO-NE Load Forecast Comm. survey results

to compare against other regional trendsto compare against other regional trends

Risks to the ForecastRisks to the Forecast

There are many. The forecast will be wrong!There are many. The forecast will be wrong!

• WeatherWeather• EconomicsEconomics• PricePrice• Data QualityData Quality• Model ErrorModel Error• Unknown and Unquantifiable – “The future doesn’t always Unknown and Unquantifiable – “The future doesn’t always

happen the way we said it would.”happen the way we said it would.”

Summary/ConclusionsSummary/Conclusions

• Load Forecasting at NU employs many of the Load Forecasting at NU employs many of the forecasting techniques covered in this classforecasting techniques covered in this class

• Each model methodology has its strengthsEach model methodology has its strengths• All are data intensive (some more than others)All are data intensive (some more than others)• Load Forecasting is not a precise scienceLoad Forecasting is not a precise science• Experience and judgment are criticalExperience and judgment are critical• The forecast will be wrongThe forecast will be wrong• Need to develop plans to manage the imprecisionNeed to develop plans to manage the imprecision