Upload
vuongtram
View
215
Download
2
Embed Size (px)
Citation preview
Webinar:UsingMachineLearningand DataAnalysistoImproveCustomer AcquisiEonandMarkeEngin ResidenEalSolar
July20th, 2017
LogisQcs
• ParQcipantsarejoinedinlisten-onlymode. • UsetheQ&ApaneltoaskquesQonsduringthewebinar.We
willhaveQmeforQ&Aattheveryend. Ø ToaskaquesQon:
o ClickQ&Aboxinthegotowebinartoolbar o TypeyourquesQonintheQ&Abox
• ForanytechnicaldifficulQesduringthewebinar, contactthe GoToWebinars HelpDeskat888.259.3826forassistance.
NATIONAL RENEWABLE ENERGY LABORATORY 2
NATIONAL RENEWABLE ENERGY LABORATORY 3
• KiranLakkaraju–SandiaNa0onalLaboratories• Yevgeniy(Eugene)Vorobeychik–VanderbiltUniversity• MichaelRossol–Na0onalRenewableEnergyLaboratory
Presenters
Photos placed in horizontal position with even amount of white space
between photos and header
Sandia National Laboratories is a multi-program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000.
Promo?ngSolarTechnologyDiffusionThroughData-DrivenBehaviorModeling
Co-PI:Dr.KiranLakkaraju,SandiaNa?onalLaboratoriesCo-PI:Prof.EugeneVorobeychik,VanderbiltUniversity
Team:UniversityofPennsylvania,CaliforniaCenterforSustainableEnergy,Na?onalRenewableEnergyLaboratory,VoteSolar
adoption probability
observability
salience
frame of reference
information/framing
income/ economic outlook
adoption by friends/
neighbors
attitudes of friends/
neighbors
attitudes/beliefs
messaging, framing
immediate net costs
long-term net value
economic uncertainty
online “build your PV” tool
subsidies
gifts
Economic Variables
Psychological Variables
Social Variables Computational Predictive Model
informational website
SAND2014-18601 PE
Residen0alPhotovoltaicsandSocCosts§ DOESunShotGoal:Reducecostofinstalledsolarenergy
systemsto$.06/kWhby20203.
§ Residen0alenergyuseis21%ofconsump0onintheUS(asof2013)1.
§ PVpricesgoingdownbut67%oftotalresiden0alsystempricegoingto“soccosts”2:§ Customeracquisi?on~9.2%§ Installa0onlabor~10.5%§ Supplychaincosts~11.7%
§ Ifwecanreducecustomeracquisi0oncosts,wereducecostsofPV.
51. http://www.eia.gov/beta/MER/?tbl=T02.01#/?f=A&start=1949&end=2013&charted=3-6-9-12
2. Friedman, Ardani, Feldman, Citron, Margoli, Zuboy, “Benchmarking Non-Hardware Balance-of-System (Soft) Costs for U.S. Photovoltaic Systems, Using a Bottom-Up Approach and Installer Survey – Second Edition”
3. http://energy.gov/eere/sunshot/mission
OverarchingResearchQues0ons
6
How can we determine residential solar PV adoption tendencies and trends, at the individual and
aggregate level?
GOAL: Develop a model to test and identify incentive policy structures that increase adoption
7
Political Ideology
Social Influence
Climate Change
Attitudes
Behavioral Economics
Incentives
Financing Structure
Retirement
Solar Community Orgs
Group Buy Environmental Ideology
Adoption Data
Taxes
Solar Irradiance
Module type
Inverters
Factors/Variables
adoption probability
observability
salience
frame of reference
information/framing
income/ economic outlook
adoption by friends/
neighbors
attitudes of friends/
neighbors
attitudes/beliefs
messaging, framing
immediate net costs
long-term net value
economic uncertainty
online “build your PV” tool
subsidies
gifts
Economic Variables
Psychological Variables
Social Variables Computational Predictive Model
informational website
Predictive Model
Experimental Data
OverarchingResearchQues0ons
8
Agent-based model (ABM) development through data-driven machine learning
Approach:
Data gathering Predictive Model ABM Modeling
0.00
0.01
0.02
100 200 300 400n
dens
ity
policyCSI Designbest 1x Bgtbest 2x Bgtbest 4x Bgtbest 8x Bgt
Policy Testing
How can we determine residential solar PV adoption tendencies and trends, at the individual and
aggregate level?
GOAL: Develop a model to test and identify incentive policy structures that increase adoption
9
Prof. Howard Kunreuther Ben Sigrin
Dr. Kiran Lakkaraju (PI) Prof. Eugene Vorobeychik (Co-PI)
Lead Organization
Tim Treadwell
Georgina Arreola
Prof. Arthur van Benthem
Prof. Ruben Lobel
Dr. Dena Gromet
Jesse Denver
Kevin Armstrong
Predic0veModel
10
Predictive Model (prob. of adoption)
• Predict probability of anindividual adopting at aparticular time.
• Identify important features.
AgentBasedModel(ABM)
11
ABM simulates emergent behavior (adoption)
Predictive Model
Interven0ons
12
+ Incentive Structure
+ Incentive Structure
ABM allows analysis of policy interventions
Whathavewelearnedfromthedata?
§ Re0rementspursthoughtsaboutadop0on(Survey).§ Monthlybillsavingsisanimportantfactor(Survey).§ Liberalsrespondmorefavorablytoa“reduce”electricity
usagemessage;conserva0vemorefavorablyto“increase”savingsmessage(Experiments).
§ Step-wiseincen0vestructures(liketheCaliforniaSolar
Ini0a0vestructure)donotincreaseadop0onwithincreasedbudgets.MoreusefultogiveawayfreesysteminourABMmodel(ABMsimula0ons).
13
Y E V G E N I Y ( E U G E N E ) V O R O B E Y C H I K
ALGORITHMIC MARKETING
Assistant Professor, EECS Director, Computational Economics Research Laboratory
Vanderbilt University
PROMOTION OF INNOVATION
• Four important questions:• Causality: which variables causally influence adoption of
innovation• Forecasting: how can we accurately forecast adoption
trends, both individual and aggregate• Impact: what is the impact of changes in environment/
policy variables on attitudes, adoption decisions, and trends• Optimization: how to we choose policy to optimally
promote innovation (subject to cost/budget constraints)
PROMOTION OF INNOVATION
• Econometrics, social science research• Causality: which variables causally influence adoption of
innovation• Forecasting: how can we accurately forecast adoption
trends, both individual and aggregate• Impact: what is the impact of changes in environment/
policy variables on attitudes, adoption decisions, and trends• Optimization: how to we choose policy to optimally
promote innovation (subject to cost/budget constraints)
PROMOTION OF INNOVATION
• Computer Science• Causality: which variables causally influence adoption of
innovation• Forecasting: how can we accurately forecast adoption
trends, both individual and aggregate• Impact: what is the impact of changes in environment/
policy variables on attitudes, adoption decisions, and trends• Optimization: how to we choose policy to optimally
promote innovation (subject to cost/budget constraints)
PROMOTION OF INNOVATION
• Computer Science • Causality: which variables causally influence adoption of
innovation • Forecasting: how can we accurately forecast adoption
trends, both individual and aggregate • Optimize model accuracy in predicting unseen data • Crucial: embed known causal models and controls • Constrain models to make economic sense
• Impact: what is the impact of changes in environment/policy variables on attitudes, adoption decisions, and trends
• Optimization: how to we choose policy to optimally promote innovation (subject to cost/budget constraints)
PROMOTION OF INNOVATION
• Computer Science • Causality: which variables causally influence adoption of
innovation • Forecasting: how can we accurately forecast adoption
trends, both individual and aggregate • Impact: what is the impact of changes in environment/
policy variables on attitudes, adoption decisions, and trends • Optimization: how to we choose policy to optimally
promote innovation (subject to cost/budget constraints) • Assuming that causality of policy variables is correctly captured
ALGORITHMIC PERSPECTIVE ON PROMOTION OF INNOVATION
adoption model
ALGORITHMIC PERSPECTIVE ON PROMOTION OF INNOVATION
Aggregate adoption forecast
adoption model
ALGORITHMIC PERSPECTIVE ON PROMOTION OF INNOVATION
Aggregate adoption forecast
promotion strategies
…
policy variables
adoption model
Exogenous (environment)
variables
ALGORITHMIC PERSPECTIVE ON PROMOTION OF INNOVATION
Aggregate adoption forecast
promotion strategies
…
policy variables
adoption model
Exogenous (environment)
variables
Budget constraint
ALGORITHMIC PROBLEM: FORECASTING
• Problem: • Use prior individual-level adoption data to
develop a model for forecasting individual decisions, and aggregate adoption trends
• Technical challenges: scalable model calibration, model validation at multiple scales
• Approach: • Data-driven agent-based modeling (DDABM) • Scalable calibration: using machine learning
(ensuring features are causally grounded) • Validation: cross-validation (measure
effectiveness of individual behavior); forecasting validation (split data along time dimension into calibration and validation sets)
adoption model
DDABM1
• Individual agent behavior: probability that individual (household) adopts solar PV, as a function of a collection of variables (including policy variables) • Learn from a large individual-event dataset of past
adoption decisions • Include social influence: capture interactions among
agents; leverage previous causal modeling method (Bollinger & Gillingham, 2012)
• Agent-based model: learned agent models instantiated as households, given household/demographic parameters from data
1 H Zhang, Y Vorobeychik, J Letchford, K Lakkaraju. Data-driven agent-based modeling, with application to rooftop solar adoption. International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2015.
DATA
• Adopters: adoption (incentive reservation) time, installation time, lease vs. own, system size, assessor data (property characteristic: square footage, etc), energy use data for 12 months prior to adoption (obtained from Center for Sustainable Energy)
• Owners: cost of ownership • Leasers: cost of leases (based on a subsample of
actual lease contracts) • Non-adopters: assessor data • Climate data (solar insolation/geographic location)
DDABM1
1 H Zhang, Y Vorobeychik, J Letchford, K Lakkaraju. Data-driven agent-based modeling, with application to rooftop solar adoption. International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2015.
Model captures actual adoption patterns
calibration validation
Incentives have weak impact on adoption trends
fore
ca
st
ALGORITHMIC PROBLEMS: PROMOTION
Aggregate adoption forecast
promotion strategies
…
policy variables
adoption model
Exogenous (environment)
variables
Budget constraint
OPTIMIZING INCENTIVES2
• Problem: given a budget and a time horizon, how to optimally choose incentive policy • Amount of incentive / watt • MW transition points to lower incentives
Rate($/watt)
Target (megawatt)
r*
1
m*
1
B
mi
2
r*
2
mi
3
ri
2
0.00
0.01
0.02
150 200n
dens
ity policyCSI Designbest 1x Bgtbest 2x Bgtbest 4x Bgt
2 H Zhang, Y Vorobeychik, J Letchford, K Lakkaraju. Data-driven agent-based modeling: framework and application. Journal of Autonomous Agents and Multiagent Systems, 2016.
DOOR-TO-DOOR MARKETING4
• Door-to-door marketing has been shown effective, but is also quite expensive
• How can we effectively decide where to focus marketing efforts? • Target individuals who are likely to
adopt, and who are in the best position to influence others to adopt
• Traveling to individual households is time consuming; wish to do this efficiently
4 H Zhang, Y Vorobeychik. Dynamic influence maximization under increasing returns to scale. AAAI Conference on Artificial Intelligence (AAAI), 2016.
DOOR-TO-DOOR MARKETING MODEL
31
• Walk door-to-door along a road network
• Goal: maximize # of adopters, subject to the constraint that your walk takes at most H hours
social influence
network
road network
4 H Zhang, Y Vorobeychik. Dynamic influence maximization under increasing returns to scale. AAAI Conference on Artificial Intelligence (AAAI), 2016.
ALGORITHMS FOR DOOR-TO-DOOR MARKETING
• Developed a new greedy algorithm (GCB), adding households to visit based on marginal impact on influence per unit marginal cost of visiting
• Proved that it’s near-optimal and much faster than state-of-the-art alternatives
32
0 2 4 6 8 10
020
4060
80100
Budget
Influence
GCBGRISK
0 2 4 6 8 10
010
2030
4050
Budget
Run
ning
Tim
e (s
)
GCBGRISK
4 H Zhang, Y Vorobeychik. Dynamic influence maximization under increasing returns to scale. AAAI Conference on Artificial Intelligence (AAAI), 2016.
ALGORITHMIC PROBLEM: DIVIDE BUDGET AMONG MANY MARKETING ACTIVITIES
Aggregate adoption forecast
promotion strategies
…
policy variables
adoption model
Exogenous (environment)
variables
Budget constraint
ALGORITHMIC PROBLEM: DIVIDE BUDGET AMONG MANY MARKETING ACTIVITIES
• Problem: have a collection of marketing activities, with varying costs and impact on adoption
• Approach: can formulate as an approximate multi-Knapsack problem • Developed new algorithms for efficient query strategies
when channel response is represented using simulations (e.g., ABM) Our approach
BOTTOM LINE
• Many algorithmic problems arise in the context ofinnovation diffusion• Modeling and big-data analytics of adoption decisions
(human behavior modeling)• Computational (often, combinatorial optimization) methods
for optimizing marketing decisions
• We made progress addressing some of these in thecontext of solar PV adoption, using individual-levelCSI data for San Diego county• Uncovered and addressed some fundamental
computational problems in the process
WillingnesstoAdoptSolar:Aninves?ga?onintoAddesignMichaelRossol1KimberlyWolske2,AnnikaTodd3
JamesMcCall1,andBenSigrin1July20th,20171Na0onalRenewableEnergyLaboratory2HarrisPublicPolicy,UniversityofChicago3LawrenceBerkeleyNa0onalLab
NATIONAL RENEWABLE ENERGY LABORATORY 37
Customeracquisi0onisasubstan0alpor0onofthecostofPV
Fuetal.2016U.S.SolarPhotovoltaicSystemCostBenchmark
$2.93/W
Moezzietal.In-preparaBonANon-ModelingExploraBonofResidenBalSolarPhotovoltaic(PV)AdopBonandNon-AdopBon
EnvironmentPrimarily:3-5%
Environment+Money:33%
Both,butMoneyoverEnvironment:39%
MoneyandnotEnvironment:6%
OpportunisAc:~20%
NATIONAL RENEWABLE ENERGY LABORATORY 38
H1:Consumerswillrespondmorefavorablytolossframesthangainframes
Ø Theliteraturesuggeststhatlossesloomlargerthanpoten0alequivalentgains,solossframingshouldbemoremo0va0ng
H2:AdsthatpresentthebenefitsofPVintheneartermwillbemoreeffec0vethanadsthatframethebenefitsinthelong-term
Ø Theliteraturesuggeststhatpeoplediscountthefutureandthusneartermfinancialbenefitsmaybemoreeffec0ve
H3:AdsthatpresentthebenefitsofPVintermsofnetsavings(i.e.savingsnetofthesystempayment)willbemoreeffec0vethanthoseintermsofgrosssavings(i.e.doesn’tincludesystempayment)
Ø Priorworkhassuggestedcustomersassumethesavingsaregross,whichunderstatesthepoten0albenefits
H4:AdsthatpresenthighermonetarysavingsfromPVwillbemoreeffec0vethanthosewithlowermonetarysavings.
Ø Assessesmarketpoten0alasafunc0onofthebillsavingsamount
ResearchQues0ons
NATIONAL RENEWABLE ENERGY LABORATORY 39
• Surveyso Adso Measuresusedo Demographics
• Resultso Gain/LossbyTimeFrameo Net/GrossbySavingAmounto RoleofAdSkep0cismo Roleofpre-exis0ngpsycho-socialconstructs
• Conclusions
Outline
NATIONAL RENEWABLE ENERGY LABORATORY 40
Survey1:Gain/LossbyTimeFrame
GainFaming LossFraming
SavingsarebasedonaMay2016assessmentoflikelysolarsavingsforresiden0alcustomerinCAorNYwitha5kWsystem.
CA NY
$67eachmonth $28eachmonth
$804eachyear $336eachyear
$20,100overthenext25years $8,400overthenext25years
NATIONAL RENEWABLE ENERGY LABORATORY 41
Survey2:Net/GrossbySavingAmount
Grosssavings:Ambiguous Netsavings:Unambiguous
Amountofsavings:- Low=$300peryear- Mid=$625peryear- High=$925peryear
MidsavingsarebasedonaMay2017assessmentoflikelysolarsavingsforresiden0alcustomerinMAwitha5kWsystem.
NATIONAL RENEWABLE ENERGY LABORATORY 42
• DependentVariables:o Likelihoodtorespondtotheado BasedontheadhowappealingisPV
– AppealofSolarPanels– CostbenefitofPV– FinancialImpactofPV
o Howskep0calareyouofthebenefitsofPV• Covariates:
o Psycho-social– EnvironmentalNorms– PerceivedSocialSupport– CustomerInnova0venessCNS,CUM– Homein-suitability
o Socioeconomics– Age– Gender– Educa0on– Householdsize– HomeSquareFootage– SESRung(Income)
SurveyMeasures
NATIONAL RENEWABLE ENERGY LABORATORY 43
Gain/LossandTimeFramingdidnotsignificantlyinfluencetheappealofPVorlikelihoodofrespondingtothead
Study1–Gain/LossbyTimeFrame
LossLoss
GainGain
LikelihoodtoRespondtoAd AppealofPVacerseeingtheAd
NATIONAL RENEWABLE ENERGY LABORATORY 44
HigheramountsframedasnetsavingsincreasedtheappealofPV,butdidnotincreaselikelihoodtorespondtothead
Net/GrossXAmountp=0.02
Study2–Net/GrossbySavingAmount
Net
Gross
Net
Gross
LikelihoodtoRespondtoAd AppealofPVacerseeingtheAd
NATIONAL RENEWABLE ENERGY LABORATORY 45
SurprisinglyNetadsindicategreaterskep0cism
Net/Grossp=0.018
Net
Gross
Loss
Gain
Gain/LossbyTimeFrame Net/GrossbySavingAmount
NATIONAL RENEWABLE ENERGY LABORATORY 46
Results:Psycho-SocialVariablesstronglyinfluenceWTA
LikelihoodtoRespondtoAd Skep0cismabouttheAd
NATIONAL RENEWABLE ENERGY LABORATORY 47
H1:Consumerswillrespondmorefavorablytolossframesthangainframes
Ø RespondentsrespondedsimilarlytobothgainandlossframesH2:AdsthatpresentthebenefitsofPVintheneartermwillbemoreeffec0vethanadsthatframethebenefitsinthelong-term
Ø Respondentsrespondedsimilarlyregardlessof0meframeforsavings/losses
H3:AdsthatpresentthebenefitsofPVintermsofnetsavings(i.e.savingsnetofthesystempayment)willbemoreeffec0vethanthoseintermsofgrosssavings(i.e.doesn’tincludesystempayment)
Ø NetadsresultedingreaterappealforPVbutdidnotleadtoagreaterlikelihoodtorespondtothead
H4:AdsthatpresenthighermonetarysavingsfromPVwillbemoreeffec0vethanthosewithlowermonetarysavings.
Ø Respondentsrespondedsimilarlyregardlessofadamount
Hypothesis
NATIONAL RENEWABLE ENERGY LABORATORY 48
• EffortstomakethefinancialbenefitsofPVseemmoreappealingdonotappeartostronglyinfluenceinterestinPV:o Timeframeforsavingshasliwleeffecto Gainandlossframingareequallyeffec0veo Greatersavingsdoesnotincreaselikelihoodtorespondo Providinggreaterfinancialdetailleadstogreaterskep0cism
• Insteadpre-exis0ngpsycho-socialconstructsappeartodriveinterestinPVo Beingpro-environmento HavingsocialsupporttoadoptPVo PerceivingyourhometobewellsuitedforPVo Beinginnova0ve
• Targetedmarke0nghasthepoten0alforgreatestreturno Findtheproperaudienceo Targetadstosaidaudience
Conclusions
NATIONAL RENEWABLE ENERGY LABORATORY 49
Ques0ons?
NATIONAL RENEWABLE ENERGY LABORATORY 50
Ques0ons?