Laura M. Andersen a Lars Gårn Hansen a Carsten Lynge Jensen a Frank A. Wolak b

Transcription

1 Shifting Electricity Demand Into versus Away from Hours of the Day Using Real-time Electricity Pricing and Information Provision: Experimental Evidence from Denmark * Laura M. Andersen a Lars Gårn Hansen a Carsten Lynge Jensen a Frank A. Wolak b a University of Copenhagen, Institute of Food and Resource Economics, Rolighedsvej 23, DK-1958 Frederiksberg C, Denmark (lma@ifro.ku.dk, clj@ifro.ku.dk, lgh@ifro.ku.dk). b Stanford University, Program on Energy and Sustainable Development and Department of Economics, 579 Serra Mall, Stanford CA , USA (wolak@zia.stanford.edu).

2 Motivation An increasing number of jurisdictions have implemented policies to increase significantly the contribution of renewable energy 1) California has a renewables portfolio standard (RPS) requiring 33 percent renewables by 2020 and 50 percent by ) Hawaii has 100% renewable energy goal for 2045 These goals are being met primarily with wind and solar energy. 1) A major reliability challenge is the uncertain amount of the energy these sources produce each hour. 2) Potential for significant unpredictable positive and negative supply and demand imbalances substantially changes the role of active demand-side participation. 1

3 Wind Output and Load in Denmark 7000 Figure 1(a) Consumption and wind/solar production Sum of DK-West and DK-East Markets Monday 13 to Sunday 19 January MegaWatts Mon Tue Wed Fri Sat Sun Consumption Wind production 2

4 The Duck Curve in California 3

5 Histogram of Hourly Wind and Solar Output in CA 4

6 Histogram of Hourly Combined Wind and Solar Output in CA 5

7 Implications of Intermittency Despite having more than 12,000 MW of wind and solar capacity in California in 2016, during majority of hours of the year, these units produced less than 4,000 MWh Wolak (2016) Level and Variability Trade-offs in Wind and Solar Investments: The Case of California, The Energy Journal, demonstrates very high degree of positive correlation in hourly output across CA wind locations and CA solar locations Between 2012 and 2016, both coefficient of variation and skewness in hourly total renewable energy output in CA has increased Coefficient of variation = SD(Q)/Mean(Q) Standardized Skewness = E(X E(X)) 3 /SD(Q) 3 6

8 Implications of Intermittency In a system where renewable energy share exceeds the capacity factor of the renewable resources, over-generation conditions can occur during many hours of the year Example: Average hourly demand is 100 MWh, annual renewable energy share is 33%, but renewable capacity factor is 25% 1) This requires 133 MW of renewable capacity which can produce as much as 133 MWh during certain hours of the year a) Shifting demand into hours with significant renewable energy mitigates this situation 2) Shifting demand away from hours with limited renewable energy production also important 7

9 Purpose of Paper Present results of field experiment with Danish residential consumers facing dynamic price and information signals aimed at causing them to shift their consumption 1) Into certain hours of the day 2) Away from certain hours of the day. Consumers are notified of these price and information signals through text messages to their cell phones with between 2 to 12 hours prior warning Messages tell customers to shift consumption into or away from designated time interval 8

10 Price and Information Signals Customers offered rebates on their electricity bill that depend on amount of electricity moved into or away from selected period Customers could receive a 5 percent, 20 percent, or 50 percent rebate off of the price of electricity for each KWh moved For the purely informational signals, customers were told how much greenhouse gas (GHG) emissions would be reduced as result of their load-shifting activities, but where not offered any financial compensation 9

11 Preview of Results Same marginal price signal produces a two to three times larger in absolute value estimated load shift into target hours relative to the absolute value of shift away from target hours 1) Shifting into also yields reduced consumption in the hours of the day that surround these target hours 2) Shifting away also yields slightly increased consumption in hours of the day that surround the target hours Informational signals produce qualitatively similar results 1) Absolute value of into effect is significantly larger in absolute value than the shift away effect 2) Stronger evidence that shifting into interval reduces consumption in surrounding periods than shifting away increases consumption in surrounding time periods. 10

12 Implications of Results Dynamic retail pricing plans using rebatres should reward shifting consumption into hours of the day, rather than away 1) To match hourly demand with increased supply of renewables 2) To reduce demand in surrounding hours with less renewables production This approach to pricing likely to stimulate market for distributed storage investments Limits the impact of the option to quit identified in Wolak (2010) with respect to dynamic rebate pricing programs An Experimental Comparison of Critical Peak and Hourly Pricing: The PowerCentsDC Program, on web-site 11

13 Outline of Remainder of Talk 1) Experimental design 2) Data collected 3) Empirical results 4) Placebo Tests of Valid Randomization 5) Possible Explanation for results and its implications 12

14 Design of Experiment Experiment was conducted in collaboration with electricity retailer SE in Denmark. In April 2015, an informed customers of a new SE-program called SE MOVEPOWER with one of two possible treatments 1) Customers earn a rebate if they move their power consumption into or away from particular time slot a) Information on relevant time slot would be sent though text message to customer s cell phone 2) Customers were told that SE would ensure GHG emissionsfree production of electricity equal to total amount of energy they moved 13

15 Design of Experiment Customers were randomized across seven different treatments. 1) Three treatments offered: 5%, 20%, and 50% rebate on all energy moved in accordance with the text messages 2) Four treatments promised customers that all energy moved in accordance with the text messages would be produced using GHG free energy sources a) SE committed to increase investments in GHG free energy production matching the amount of energy moved. b) Four information treatments only reflect slight differences in the wording of how this information was conveyed to the consumers. 14

16 Design of Experiment All consumers had advanced power meters which registered their hourly consumption making it possible to calculate the customer s consumption in the relevant time slots. In total, 737 customers signed up for the rebate based program and 1,065 customers signed up for the GHG-free energy program. The first text messages were sent on the 4 th of June 2015 and the experiment was terminated on 7 th of February ) Sample period used in analysis Customers were prompted a few hours in advance on the same day they were supposed to move power. 15

17 Design of Experiment Customers were notified an average of 1.2 times a week of the threehour time slots in which a rebate could be earned 1) Randomly assigned to into and away shifting Text message notified them of 1) Target time slot, 2) Whether they should move power into or away, 3) Rebate percent they would earn or the GHG-free production they would ensure by moving energy Target time slots varied randomly across the days of the week, between different 3 hour time slots Starting with 10 am to 1 pm and ending 12 am to 3 am Prior notification also varied from 2 hours to 5 hours in advance 16

18 Design of Experiment Customers received monthly showing the amount of energy they had moved up to that point in the experiment After the experiment was terminated on 7 th of February 2016 SE calculated rebates and the amount of KWH s GHG-free energy production due each customer Rebates were then paid to costumers and earned GHG-free KWH reported in connection with the following quarterly power bill. All communication with customers from the initial recruitment mail to text messages and feedback was done by SE through their mail server and text message service using their letterhead and logo. 17

19 Data Summary Statistics (1/4/15-2/7/16) Summary Statistics on Rebate Groups. 5% rebate 20% rebate 50% rebate Number of customers Average number of time slots per customer With into-treatment i With away-treatment ii With no treatment iii Average kwh consumption per 3 hours period With into-treatment With away-treatment With no treatment

20 Data Summary Statistics (1/4/15-2/7/16) Summary Statistics on Environmental Experiment Groups Group: 31 Group: 34 Group: Group: Number of customers Average number of time slots per customer With into-treatment(a) With away-treatment(b) With no treatment(c) Average kwh consumption per 3 hours period iv With into-treatment With away-treatment With no treatment

21 Estimation Procedure and Results Estimate a number of different average treatment effects for each of the three rebate groups and the four information provision groups. Treatment effects are estimated using diff-in-diff estimator Customers in each of the three rebate treatment groups are randomly assigned to receive treatment across and within days. Customers in each of the four information treatment groups are randomly assigned to receive treatment across and within days. Customers in each sample not experiencing a treatment event in that time interval and day serve as the control group used to estimate the treatment effect 20

22 Variable Definitions Day is divided into 9 time periods 3 am to 6 am, 6 am to 7 am, 7 am to 10 am, 10 am to 1 pm, 1 pm to 3 pm, 3 pm to 6 pm, 6 pm to 9 pm, 9 pm to 12 am, and 12 am 3 am. Treatment events for both the rebate and informational samples were only declared during the 3-hour time periods. 21

23 Variable Definitions For each rebate group we define six indicators, three for the intotreatment and three for away-treatment. 1) Away(r,i,t,d) is equal to 1 for rebate level r (r = 5 percent, 20 percent and 50 percent), if customer i in time period t, of day d received an away notification for time period and day and is equal to zero for all other time periods in the sample. 2) BeforeAway(r,i,t,d) is equal to 1 for all time periods after an away notification was sent to consumer i with rebate level r and before the actual away time period occurred for this customer and equal to zero for all other time periods in the sample 3) AfterAway(r,i,t,d) is equal to 1 for as many hours after the away event as the BeforeAway(r,i,t,d) variable was equal to 1 for the same away event and equal to zero for all other time periods in the sample. 22

24 Variable Definitions Three analogous variables are defined for the into events. 1) Into(r,i,t,d) is equal to 1 for rebate level r if customer i in time period t of dy d received an into notification for that time period and day and equal to zero for all other time periods in the sample 2) BeforeInto(r,i,t,d) is equal to 1 for all time periods after an into notification was sent to consumer i with rebate level r and before the actual into time period occur for this customer and equal to zero for all other time periods in the sample. 3) AfterInto(r,i,t,d) is equal to 1 for as many hours after the into event as the BeforeInto(r,i,t,d) variable was equal to 1 for the same into event and equal to zero for all other time periods in the sample. 23

25 Variable Definitions and Models Estimated y(r,i,t,d) = natural logarithm of electricity consumption in kilowatthours of customer i facing rebate level r during period t of day d. We estimate the following regression for each of the three samples of rebate customers: y(r,i,t,d) = μ(t) + ν(i) + η(d) + β 1 BeforeInto(r,i,t,d) + β 2 Into(r,i,t,d) + β 3 AfterInto(r,i,t,d) + α 1 BeforeAway(r,i,t,d) + α 2 Away(r,i,t,d) + α 3 AfterAway(r,i,t,d) + ε(r,i,t,d) where the μ(t) (t=1,2,...,9) are period of day fixed effects, the ν(i) (i=1,2,,i) are customer fixed effects, the η(d) (d=1,2,..,d) are day of sample fixed effects, and the ε(r,i,t,d) are mean zero regression disturbances. 24

26 Regression Results Results for 5 Percent, 20 Percent and 50 Percent Rebate Levels Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t 5 % Rebate 20 % Rebate 50 % Rebate Regressor BeforeInto (0.0033) (0.0045) (0.0054) Into (0.0057) (0.0077) (0.0099) AfterInto (0.0031) (0.0043) (0.0051) BeforeAway (0.0049) (0.0071) (0,0085) Away (0.0075) (0.0108) (0.0130) AfterAway # of Observations (0.0045) (0.0064) (0.0077) 694, , ,663 Standard errors computed using the heteroscedasticity and autocorrelation-consistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. 25

27 Results to Highlight Uniformly two to three times larger in absolute value coefficient on Into(r,i,t,d) versus Away(r,i,t,d). 1) The into average treatment effect ranges from a roughly 8 percent to 13 percent increase in consumption during the treatment period, and is significantly larger for the 50 percent rebate relative to the 5 percent and 20 percent rebate level. 2) The away average treatment effect is between 3 and 4.5 percent for all rebate levels, with the highest percentage reduction occurring for the 20 percent rebate level. Both before and after an into event consumption is lower than the control group 1) Results are very encouraging for using into treatments to achieve targeted demand reductions as well as a targeted demand increases. 26

28 Pooled Regression Results To investigate whether the imprecise results for the BeforeAway and AfterAway might be due a sample size issue, we estimate a pooled version of the model 1) Imposes the restriction that all three rebate groups have the same time period in the day fixed effects and the same day-ofsample fixed effects We estimate pooled regression across the three rebate groups: y(r,i,t,d) = μ(t) + ν(i) + η(d) + r=5,20,50 [β 1r BeforeInto(r,i,t,d) + β 2r Into(r,i,t,d) + β 3r AfterInto(r,i,t,d)+ α 1r BeforeAway(r,i,t,d) + α 2r Away(r,i,t,d) + α 3r AfterAway(r,i,t,d) + ε(r,i,t,d) 27

29 Pooled results for 5 percent, 20 percent and 50 percent rebate levels Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t 5 % Rebate 20 % Rebate 50 % Rebate Regressor BeforeInto (0.0033) (0.0044) (0.0053) Into (0.0056) (0.0077) (0.0098) AfterInto (0.0031) (0.0042) (0.0051) BeforeAway (0.0049) (0.0071) (0.0084) Away (0.0075) (0.0108) (0.0129) AfterAway (0.0045) (0.0063) (0.0077) 1,336,996 # of Observations Standard errors computed using the heteroscedasticity and autocorrelation-consistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. 28

30 Results to Highlight The major change in the results from individual regressions is the larger in absolute value coefficients on Into(r,i,t,d) for the 20 and 50 percent rebates level and the smaller in absolute value coefficient on Away(r,i,t,d) for the 50 percent rebate level. For the same rebate percentage, significantly larger in absolute value treatment effects hold for the into intervention versus the away intervention 29

31 Informational Treatment Similar regressions are run for informational treatments, but all informational treatments are pooled Same qualitative conclusions from the results for rebates Significantly larger in absolute value treatment effects hold for the into intervention versus the away intervention 30

32 Pooled Estimates of Impact of Four Informational Treatments Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t Group 31 Group 34 Group 35 Group 36 Regressors BeforeInto (0.0033) (0.0037) (0.0054) (0.0061) Into (0.0057) (0.0061) (0.0091) (0.0107) AfterInto (0.0031) (0.0035) (0.0050) (0.0057) BeforeAway (0.0051) (0.0056) (0.0082) (0.0094) Away (0.0078) (0.0087) (0.0127) (0.0146) Afterway (0.0045) (0.0051) (0.0074) (0.0086) 1,703,242 # of Observations Standard errors computed using the heteroscedasticity and autocorrelation-consistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. 31

33 Restricted Estimates of Impact of Informational Treatments Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t Regressor BeforeInto (0.0021) Into (0.0036) AfterInto (0.0020) BeforeAway (0.0032) Away (0.0050) Afterway (0.0029) # of Observations 1,703,242 Standard errors computed using the heteroscedasticity and autocorrelationconsistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. 32

34 Testing the Validity of the Randomization of Interventions Several placebo regressions were run to investigate whether our into and away interventions actually cause the consumption changes that we estimate in the previous section. Define Two Indicator Variables: 1) IntoP(r,i,t,d) equals 1 in time period t of day d if this time period is immediately before notification of an into intervention for customer i with rebate level r and zero in all other time periods 2) AwayP(r,i,t,d) equals 1 in time period t of day if this time period is immediately before notification of an away intervention for customer i with rebate level r and zero in all other time periods. 33

35 Testing the Validity of the Randomization of Interventions For each rebate level sample and the pooled rebate sample, we estimate the following regression: y(r,i,t,d) = μ(t) + ν(i) + η(d) + βintop(r,i,t,d) + αawayp(r,i,t,d) + ε(r,i,t,d) For each regression we would not expect either α or β to be nonzero because customers have no economic or informational incentive to shift their consumption into or away from time periods when either IntroP or AwayP is equal to 1. 34

36 Placebo Estimates of Impact of Rebate Treatments Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t 5 Percent Rebate 20 Percent Rebate 50 Percent Rebate Pooled Sample Regressor IntoP (0.0082) (0.0119) (0.0164) (0.0063) AwayP (0.0114) (0.0142) (0.0163) (0.0079) # of Obs 697, , ,734 1,342,618 Standard errors computed using the heteroscedasticity and autocorrelationconsistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. 35

37 Placebo Estimates of Impact of Informational Treatments Dependent Variable is Natural Logarithm of Customer i s Consumption in Time Period t Regressor IntoP (0.0059) AwayP (0.0071) # of Observations 1, Standard errors computed using the heteroscedasticity and autocorrelationconsistent covariance matrix for two-way panel data models presented in Arellano (1987) are in parentheses below coefficient estimates. Results are consistent with the into and away consumption shifting estimates presented in the previous section being caused by our rebate and informational treatments. 36

38 Cost-Effectiveness of the Shifting Into and Shifting Away Total KWh shifted for each customer is computed as follows. For each into event and rebate level: MoveI(r,i,t,d)= (1 - exp(-b r ))C(r,i,t,d) where b r is the estimate of β 2, the average treatment effect from the pooled estimates and C(r,i,t,d) is the customer s actual consumption during period t of day d. For each away event and rebate level, we estimate the total amount of energy moved during that time period as MoveA(r,i,t,d) = (exp(-b r ) 1)C(r,i,t,d) We compute the sum of MoveI and MoveA over all interventions that customer i experienced during the experiment and divide that magnitude into the total amount of rebates in Kroner paid for customer i. 37

39 Histogram of Average Rebate Paid Per KWh Moved 38

40 Histogram of Customer Level Average KWh Shifted per Into Intervention 39

41 Histogram of Customer-Level Average KWh Shifted per Away Intervention 40

42 Summary of Results More that 50 percent of the customers have values less than 0.5 Kroner/KWh, which is roughly 0.07 Dollar/KWh. MWh Reductions from Into and Away Events The values range from zero to 0.8 KWh for into intervention with a median value of 0.2 KWh for into intervention. For the away interventions the numbers range from zero to 0.2 KWh per away intervention with a median of 0.05 KWh per away intervention. Four times larger average amount shifted for into interventions suggests that this form of consumption-shifting is likely to be a far more cost-effective way to maintain system balance in regions with significant intermittent renewable generation capacity. 41

43 Explaining Results (The Option to Quit) Suppose that customer has two possible realizations of demand when an event period is called D L (p) = low demand (no guests at house) D H (p) = high demand (having guests at house) Q = actual consumption of electricity p N = normal price of electricity r = rebate level Q R = reference level relative to which rebates are issued Assume for simplicity that Q R is the same for into and away interventions 42

44 Explaining Results (The Option to Quit) For away events the marginal price of additional consumption is (p N + r) for Q < Q R and p N otherwise For into events marginal price of additional consumption is (p N - r) for Q > Q R and p N otherwise The option to quit under a rebate scheme discussion in Wolak (2011) can explain the higher in absolute average treatment effect of an into event relative to the absolute value of the average treatment effect of an away event 1) Customer finds it optimal not to response to away event in high demand state because its optimal price response does not earn a rebate 43

45 INTO 44

46 AWAY 45

47 Putting INTO and AWAY Results Together Both Outcomes Support ATE(INTO) > ATE(AWAY) 46

48 Conclusions Experiment suggest an alternative more cost-effective rebate mechanism for active participation of the final consumers in managing the real-time supply and demand balance in regions with significant intermittent renewable generation For the same rebate percentage for load-shifting into a time period induced a two to three times larger percent increase in demand than that rebate induced for load-shifting away from that time period. 1) Better addresses option-to-quit problem with rebates A significant amount of the energy that shifted into the time period also resulted in reductions in consumption during time periods before and after this time period. 47

49 Conclusions For purely informational interventions produced analogous results: 1) Significantly larger in absolute value average load-shifting into time periods relative to shifting away from time periods and evidence that load-shifting into a time period led to lower consumption during neighboring time periods 2) Load-shifting away from a time period did not consistently lead to increases in consumption in neighboring periods 48

50 Questions? More Information at 49