Residential Customer Response to RealTime Pricing: The Anaheim CriticalPeak Pricing Experiment

1
Residential Customer Response to Real-Time
Pricing The Anaheim Critical-Peak Pricing
Experiment

• Frank A. Wolak
• Department of Economics
• Stanford University
• Stanford, CA 94305-6072
• wolak_at_zia.stanford.edu
• http//www.stanford.edu/wolak

2
Outline of Talk

• Description of experiment design
• Assessing validity of experimental design
• Measurement framework employed
• Treatment effect of CPP event
• Sensitivity of estimation results to assumptions
• What does treatment effect measure?
• Reference level inflation
• Setting reference level for rebates for budget
  balance
• CPP with rebate as transition to default
  real-time pricing of electricity to residential
  consumers

3
Anaheim CPP Experiment

• During the summer of 2005, the City of Anaheim
  Public Utilities (APU) ran a Critical Peak
  Pricing (CPP) experiment
• During late 2004, a random sample of APU
  residential customers were selected to
  participate in experiment
• Customers in this sample were randomly assigned
  to the control and treatment groups
• Control customers were not notified of this
  selection but simply had interval meters
  installed at their dwelling
• Treatment group customers first received a letter
  notifying them that they had been selected to
  participate in CPP program and were asked to
  return a reply form with their phone number
  and/or e-mail address
• Follow-up phone calls to sign-up those that did
  not respond to mailing
• Follow-up mailing to recruit those who could not
  be contacted by phone
• Final result--Very little attrition from randomly
  selected treatment group
• Process ultimately resulted in 52 control
  customers and 71 treatment customers, or 123
  total customers

4
Anaheim CPP Experiment

• All customers (treatment and control) paid a
  fixed retail price of 6.75 cents/kWh up to their
  monthly baseline of 240 KWh
• Monthly consumption beyond 240 KWh baseline
  charged at 11.02 cents/kWh
• Customers in treatment group were subject to a
  maximum of 12 CPP days for experiment period
• Day-ahead notification of CCP days via telephone
  or e-mail (depending on customers choice on
  reply card)

5
Anaheim CPP Experiment

• CCPs days are required to be on weekdays that are
  not holidays
• Consumption below reference level during peak
  period (noon to 6 pm) of CPP days eligible for
  refund of 35 cents/KWh
• Consumers receive a rebate if their average
  consumption during peak periods of CPP days is
  less than their reference peak period consumption
  
• Rebate on day d max(0,(q(ref)
  q(peak,d)))p(rebate)
• Rebate implies that customers guaranteed not to
  pay more than they would have under control
  tariff
• Reference peak period consumption is customers
  typical peak period consumption
• Defined as average peak period consumption during
  three highest non-CPP days (excluding weekends
  and holidays during experiment)
• All CPP-eligible days that were not CPP-days
  during experiment

6
Dataset Used in Analysis

• Daily Peak and Off-peak period consumption for
  123 locations Peak periodnoon to 6 pm
• Peak(i,d) Peak period consumption for location
  i on day d
• Off-periodall other hours of the day
• OffPeak(i,d) Off-Peak period consumption for
  location i on day d
• Temp(d) Maximum daily temperature at Fullerton
  Airport for day d
• Day(d) Indicator for whether day d1,,136 (all
  days during sample period)
• LOC(i) Indicator for location i, i1,,123
• Treat(i) Indicator for whether location i is in
  treatment group
• CCP(d) Indicator for whether day d is a
  critical peak day

7
Pre-Treatment Period Comparison

• Meters installed for all customers in experiment
  before June 1, 2005 start date of experiment
• Consumption recorded at 15-minute intervals
  throughout the day for customers in both groups
• Comparison of pre-treatment 15-minute means to
  assess randomness of selection of customers into
  experiment and their assignment to treatment and
  control groups

8
Pre-Treatment Period Comparison

• For virtually all 15-minute periods, 95 percent
  confidence interval on mean difference in
  pre-experiment period consumption by treatment
  and controls groups contains zero
• ConclusionNo evidence of non-random selection
  into experiment or subsequently into treatment
  versus control groups

9
Measuring Price Response

• Two models estimated for peak period
• Average peak period treatment effect
• ln(Peak(i,d)) aCCP(d)Treat(i) ?d di eid
• di location-specific fixed effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• eid observable mean zero stochastic
  disturbance
• Temperature dependent peak period treatment
  effect
• ln(Peak(i,d)) aCCP(d)Treat(i)
  ?CPP(d)Treat(i)TEMP(d) ?d µi ?id
• mi location specific fixed-effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• ?id observable mean zero stochastic disturbance

10
Measuring Price Response

• Two models estimated for off-peak period
• Average off-peak period treatment effect
• ln(Off-Peak(i,d)) aCCP(d)Treat(i) ?d di
  eid
• di location-specific fixed effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• eid observable mean zero stochastic
  disturbance
• Temperature dependent peak period treatment
  effect
• ln(Off-Peak(i,d)) aCCP(d)Treat(i)
  ?CPP(d)Treat(i)TEMP(d) ?d µi ?id
• mi location specific fixed-effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• ?id observable mean zero stochastic disturbance

11
Estimation Results
12
Estimation Results
Arrellano (1987) covariance matrix used,
Estimates computed using Cochrane-Orcutt
procedure assuming AR(2) errors. All
regressions include 135 day-of-sample fixed
effects
13
Temperature Dependent Treatment Effects
14
Dynamics of Price Response

• Examine if substitution across days occurred as a
  result of CCP days
• Include lagged value of CPP(d)Treat(i)
• ln(Peak(i,d)) a1CCP(d)Treat(i)
  a1CCP(d-1)Treat(i) ?d di eid
• di location-specific fixed effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• eid observable mean zero stochastic
  disturbance
• Include lagged value of CPP(d)Treat(i)
• ln(Off-Peak(i,d)) aCCP(d)Treat(i)
  ?CPP(d-1)Treat(i) ?d µi ?id
• mi location specific fixed-effect (controls for
  persistent differences in consumption across
  locations)
• ?d day-specific fixed effect (controls for
  persistent differences in consumption across days
  in sample)
• ?id observable mean zero stochastic disturbance
• Same regression with lead value of CPP(d)Treat(i)

15
Estimation Results
No evidence of lagged or lead effects (model with
CPP(d1)Treat(i) instead of CPP(d-1)Treat(i))
of CCP events for treatment group
Arrellano (1987) covariance matrix used.
16
Summary of Results

• Load-reduction effect--Peak period consumption of
  treated group approximately 13 lower than
  consumption of control group during CCP days
• Controlling for all fixed differences across
  locations, and fixed differences across days
• Load-reduction effectEvidence of larger
  consumption reduction in higher temperature days
• Five degree temperature increase implies 4
  percentage point increase in the consumption
  reduction of treated group versus control group
• Little evidence of load shifting to off-peak
  periods
• No statistically significant difference in
  treatment versus control group mean consumption
  during off-peak periods on CPP days
• No statistically significant difference in
  treatment versus control group mean consumption
  during peak and off-peak periods in day before or
  day after CPP day

17
Customer-Level Heterogeneity in Treatment Group
18
Customer-Level Heterogeneity in Control Group
19
Total Expenditure Below Baseline of 240
KWhTreatment Group
20
Total Expenditure Below Baseline of 240
KWhControl Customers
21
Total Expenditure Above Baseline of 240
KWhControl Customers
22
Total Expenditure Above Baseline of 240
KWhControl Customers
23
Rebates Received

• All customers in treatment group benefited from
  program
• Some benefited enormously--One customer was paid
  rebates equal to 40 of its bill over the
  experiment period
• Why did some customers benefit so much more than
  others?
• Potential for reference level inflationIncrease
  consumption during non-CPP day peak periods that
  are eligible to be CPP days to increase reference
  level
• Reference level set too high so that rebates
  would be paid even it customer did not respond to
  CPP day
• Answer first question by comparing mean
  consumption of treatment and control groups
  during peak periods on non-CPP days that are
  eligible to be CPP days and therefore enter into
  reference level calculation
• Answer second question by comparing rebates that
  control group (which had no incentive to increase
  reference level or reduce consumption during CPP
  days) would have received to those received by
  treatment group

24
Total Rebates Received for 12 CPP DaysTreatment
Customers
25
Total Rebates for 12 CPP Days that Would Been
Received by Control Customers
26
Total Rebates Received Divided by Total
BillTreatment Customers
27
Total Rebates that Would Have Been Paid Divided
by Total Bill (Control Customers)
28
Reference Level Inflation

• Treatment customers can influence reference level
  relative to which refunds are issued by how they
  consume during CPP-eligible days that are not CPP
  days
• QuestionWhat impact did process used to set
  reference level have on magnitude of treatment
  effect?
• Compare mean consumption of treatment and control
  groups during days used to determine a customers
  reference peak period consumption relative to
  which rebates were computed

29
Reference Level Inflation

• Treatment customers have 7 percent higher
  consumption than control in non-CPP days that are
  eligible to be CPP days
• Approximately half of estimated treatment
  effect of CPP event due to inflation in reference
  level by treatment group
• Treatment customers have 14 percent higher
  consumption than control group during non-CPP
  days that are eligible to be CPP days
• Major challenge to formulating CPP pricing
  program with rebate is how to set reference level
• Second challenge--Paying for consumption
  reductions relative to reference level that are
  far greater than actual consumption reductions
  due to a CCP event

30
Total KWh Reduction Predicted by Treatment Effect
for 12 CPP Days (Treatment Customers)
31
Total KWh Paid 35 cents/KWh Rebate Over All 12
CPP Days (Treatment Customers)
32
KWh Reduction Predicted by Treatment Effect for
12 CPP Days (Control Customers)
33
KWh that Would Have Been Paid 35 cents/KWh Rebate
over All 12 CPP Days (Control Customers)
34
Setting Reference Level

• Pay treatment customers for 7 times more KWh
  than treatment effect says they reduced
  consumption in CPP peak period
• Would have paid control customers for 6 times
  more KWh than their predicted decrease during CPP
  period
• Setting reference level too high can make it very
  hard for CPP with rebate to satisfy cost/benefit
  test
• Savings from wholesale energy purchase costs
  greater than total rebates paid plus other costs
  program

35
Optimal Reference Level

• If goal of CPP pricing is to create predictable
  and substantial load reduction on day-ahead
  basis, some reference level inflation may be
  optimal
• Customers consume more in other peak periods to
  be able to predictably reduce their consumption
  by a substantial amount during CPP periods
• Predictable and sizeable load reduction on a
  day-ahead basis allows retailers to discipline
  unilateral exercise of market power by suppliers
  in short-term energy and ancillary services
  markets
• Predictable and sizeable load reduction on a
  day-ahead basis can significantly enhance system
  reliability
• Trade-off between process used to set reference
  level and magnitude of rebate paid during peak
  hours of CPP day
• Large rebate payment with low reference level
  may cause customers to give up on reducing demand
  during certain CPP periods, which reduces
  predictability and size of response
• Estimating dynamic model of electricity
  consumption outlined in paper can provide useful
  input to answering these questions

36
Conclusions

• Load reduction due to CPP event confined to load
  period in which event occurs
• Load reduction of close to 13 confined to peak
  periods on CPP days
• Evidence of larger percentage load reductions in
  peak periods on higher temperature CPP days
• Strong evidence of reference price inflation by
  treatment customers
• Strong evidence that reference level set too high
  for many customers
• Neither of the above two results imply that CPP
  with rebate cannot provide predictable and
  substantial load reduction that satisfies
  cost/benefit test
• Some reference level inflation and some payment
  for load reductions beyond those induced by CPP
  event may be optimal
• CPP with rebate consistent with provisions of
  AB1X which requires no customer pay more than
  fixed rate customers
• Can help to demonstrate economic and reliability
  value of real-time pricing to load while still
  being consistent with AB1X

37

• Questions/Comments
• For more information
• http//wolak.stanford.edu/wolak

38
Limited Benefits of Restructuring in US Without
Involving Demand

• US has privately-owned, profit-maximizing firms
  facing cost-of-service price regulation or
  incentive regulation plan
• Detailed prudence review of investment
• Hard to argue there are large deviations from
  minimum cost production
• Vertically integrated ownership and centralized
  dispatch should be able to improve on bid-based
  dispatch on true production cost basis

39
Markets use prices to allocate scarce resources

• Competitive market should be able to get by with
  lower level of capacity and serve same customers
• This implies lower capacity costs for market at
  large
• If dispatch costs are close to the same, then
  average price in competitive market should be
  less than average price in regulated market
• A necessary condition for this to occur is a
  sufficient number of price-responsive consumers

40
Optimal Capacity Choice Under Regulation versus
Competition
Kreg gtgt Kcomp
41
Example--US Airline Industry

• Load Factors (Seats Filled)/(Seats Total),
• In regulated regime highest load factors
  approximately 55 in 1976
• Currently Load Factors are close to 75
• This increased capacity utilization rate allows
  real average fare per passenger-mile to be
  significantly less than under regulated regime
• Regime works because of large number of
  sophisticated price-responsive consumers.

42
Even Residential Consumers Can Respond
Weekly Consumption Monday to Sunday
43
Even Residential Consumers Can Respond
Weekly Consumption Monday to Sunday
44
Even Residential Consumers Can Respond
Weekly Consumption Monday to Sunday
45
(No Transcript)