<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>Willingness to delay charging of electric vehicles</title></titleStmt>
			<publicationStmt>
				<publisher></publisher>
				<date>2020 Summer</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10209596</idno>
					<idno type="doi"></idno>
					<title level='j'>25th Annual Conference of the European Association Environmental and Resource Economists</title>
<idno></idno>
<biblScope unit="volume"></biblScope>
<biblScope unit="issue"></biblScope>					

					<author>R.A. Daziano</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[Electrification of vehicles is becoming one of the main avenues for decarbonization of the transportation market. To reduce stress on the energy grid, large-scale charging will require optimal scheduling of when electricity is delivered to vehicles. Coordinated electric-vehicle charging can produce optimal, flattened loads that would improve reliability of the power system as well as reduce system costs and emissions. However, a challenge for successful introduction of coordinated deadline-scheduling of residential charging comes from the demand side: customers would need to be willing both to defer charging their vehicles and to accept less than a 100% target for battery charge. Within a coordinated electric-vehicle charging pilot run by the local utility in upstate New York, this study analyzes the necessary incentives for customers to accept giving up control of when charging of their vehicles takes place. Using data from a choice experiment implemented in an online survey of electric-vehicle owners and lessees in upstate New York (N=462), we make inference on the willingness to pay for features of hypothetical coordinated electric-vehicle charging programs. To address unobserved preference heterogeneity, we apply Variational Bayes (VB) inference to a mixed logit model. Stochastic variational inference has recently emerged as a fast and computationally-efficient alternative to Markov chain Monte Carlo (MCMC) methods for scalable Bayesian estimation of discrete choice models. Our results show that individuals negatively perceive the duration of the timeframe in which the energy provider would be allowed to defer charging, even though both the desired target for battery charge and deadline would be respected. This negative monetary valuation is evidenced by an expected average reduction in the annual fee of joining the charging program of $2.64 per hour of control yielded to the energy provider. Our results also provide evidence of substantial heterogeneity in preferences. For example, the 25% quantile of the posterior distribution of the mean of the willingness to accept an additional hour of control yielded to the utility is $5.16. However, the negative valuation of the timeframe for deferring charging is compensated by positive valuation of emission savings coming from switching charging to periods of the day with a higher proportion of generation from renewable sources. Customers also positively valued discounts in the price of energy delivery.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="1">Introduction</head><p>Electrification of vehicles is becoming one of the main avenues for decarbonization of the transportation market. Even though there are clear environmental benefits of renewable-based electromobility, large-scale charging from high penetration of electric vehicles (EVs) will require optimal scheduling of when electricity is delivered to vehicles <ref type="bibr">(Gonzlez-Garrido et al., 2019;</ref><ref type="bibr">Calearo et al., 2019;</ref><ref type="bibr">Andersen et al., 2018;</ref><ref type="bibr">Bitar and Xu, 2017;</ref><ref type="bibr">Arif et al., 2016)</ref>. Optimal scheduling of electric-vehicle charging has the potential to reduce load variance. In fact, coordinating EVs to charge at times when fewer people require electricity can effectively prevent stress on the power grid by reducing peak loads. Smart EV charging and the resulting flattened loads can improve reliability of the power system as well as reduce system costs and emissions. From the demand side of coordinated EV charging, residential customers would need to be willing both to delay charging their vehicles and to accept less than a 100% target for battery charge.</p><p>OptimizEV is a pilot program NYSEG -the local electricity and gas provider in upstate New York -is running within the Energy Smart Community of Tompkins County, NY to precisely analyze residential optimal scheduling of the charging of electric vehicles. Following an algorithm developed by Cornell researchers, OptimizEV: 1) determines exactly when to charge an EV within both a timeframe and target charge specified by the customer, 2) offers a discount based on how long an EV is left plugged in, and 3) ensures the EV is ready to go when needed.</p><p>From a perspective of demand-side dynamics <ref type="bibr">(Chakraborty et al., 2019;</ref><ref type="bibr">Daina, 2018)</ref>, there has been increasing interest in the literature regarding modeling EV charging behavior (for reviews, see <ref type="bibr">Hardman et al., 2018;</ref><ref type="bibr">Daina et al., 2017b)</ref>. Within this avenue of research, random utility maximization models have been used to explore response to smart EV charging services. For example, <ref type="bibr">Daina et al. (2017a)</ref> built a model for joint decisions of EV charging and use (activity-travel) within the context of simplified (two-period) time-of-use pricing of electricity. Via a choice experiment in which respondents chose their target battery level and deadline to achieve the desired target given specific travel needs for the day, the study provided evidence of large heterogeneity in charging preferences. This observation about large variability in behavior is also reported in previous work <ref type="bibr">(Yang et al., 2016;</ref><ref type="bibr">Zoepf et al., 2013;</ref><ref type="bibr">Franke and Krems, 2013)</ref>. In applied microeconometrics, mixed logit choice models <ref type="bibr">(McFadden and Train, 2000)</ref> address unobserved heterogeneity in preferences through a parametric approach.</p><p>In this paper, our focus in on modeling choice of EV charging programs that implement optimal scheduling of EV charging (cf. <ref type="bibr">Richter and Pollitt, 2018)</ref>. In particular, we are interested in determining behavioral response to the idea of giving up control of EV charging by letting the electric utility to decide when to deliver electricity within a given time window and a pre-specified state of charge and deadline. In this regard, this work is related to how customers respond to terms of energy contracts including pricing, which in the literature there is also evidence of heterogeneous behavior <ref type="bibr">(Richter and Pollitt, 2018)</ref>.</p><p>To address and measure preference heterogeneity in the willingness to delay charging of electric vehicles, we propose a mixed logit model, which is fitted by Variational Bayesian (VB) methods using data from a choice experiment designed for this study. Bayes estimation is an alternative approach to the more traditional maximum likelihood estimator, with many associated benefits including direct inference on the full posterior distribution of individual-specific preference parameters that represent how tastes vary. Whereas Bayes estimators are typically simulated using Markov chain Monte Carlo (MCMC; <ref type="bibr">Rossi (2015)</ref>; <ref type="bibr">Rossi et al. (2012)</ref>) posterior sampling, Variational Bayes methods have emerged as a scalable alternative to MCMC in the domains of probabilistic machine learning and computational statistics <ref type="bibr">(Blei et al., 2017;</ref><ref type="bibr">Jordan et al., 1999;</ref><ref type="bibr">Ormerod and Wand, 2010)</ref>. Variational Bayes is implemented as an optimization problem rather than a sampling problem. The objective of stochastic variational inference is to find an approximate parametric distribution of the model parameters such that the probability distance (typically measured in terms of the Kullback-Leibler divergence) between the exact posterior distribution and the variational distribution is minimal. A key challenge in the application of VB to posterior inference in discrete choice models is that the expectation of the logarithm of the choice probabilities -i.e., the expectation of the log-sum of exponentials -lacks a closed form, due to the lack of a general conjugate prior. Whereas VB posterior inference for mixed logit models has been analyzed in the literature <ref type="bibr">(Braun and McAuliffe, 2010;</ref><ref type="bibr">Depraetere and Vandebroek, 2017;</ref><ref type="bibr">Tan, 2017)</ref>, recently in <ref type="bibr">Bansal et al. (2020)</ref> we resolved major research gaps in terms of parameter recovery, finite-sample properties, and extensions to more general utility specifications and representations of unobserved preference heterogeneity. Across the VB implemtations in <ref type="bibr">Bansal et al. (2020)</ref>, VB-NCVMP-&#8710; <ref type="bibr">(Depraetere and Vandebroek, 2017;</ref><ref type="bibr">Tan, 2017)</ref> is on average between 1.7 to 16.2 times faster than MCMC and MSLE, while performing nearly as well at prediction and parameter recovery. Thus, in this paper we explore preference heterogeneity in the responses to coordinated EV charging using VB-NCVMP-&#8710; inference.</p><p>The rest of the paper is organized as follows. Section 2 reviews the microdata, including details of the choice experiment that was implemented in the online OptimizEV survey. Section 3 provides a short description of the VB-NCVMP-&#8710; Bayes estimator that use to make inference on the degree of preference heterogeneity in preferences. Full details on the derivation of the estimator are provided in <ref type="bibr">Bansal et al. (2020)</ref>. Section 4 discusses point and interval estimates of marginal utilities and willingness to pay measures, with an emphasis of the extent of preference heterogeneity that is evidenced from the data. Finally, section 5 concludes. Another goal of the survey was to inform design of the user interface of the mobile app to communicate with the OptimizEV smart chargers. In fact, the next subsection describes a choice experiment Respondents to the survey (N=462) either own (69%) or lease (27%) mostly plug-in hybrids (PHEVs), with pure battery electric vehicles (BEVs) representing around one third of the sample (Fig. <ref type="figure">2</ref>). Of the 78% of respondents who typically leave their EVs plugged in until it is fully charged, 60% use a Level 1 charger at home. Other characteristics of the sample include: 55% have a graduate or professional degree; 90% live in a detached, single family home; 24% have onsite solar at home; 66% are employed full time; and 24% are retired. </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2">Data</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.1">The OptimizEV survey</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.2">Choice experiment</head><p>To elicit customer preferences around smart residential EV charging and deadline scheduling, a discrete choice experiment was designed mimicking some of the decisions expected from individuals participating in the OptimizEV program.</p><p>The choice experiment was presented to respondents as the electricity provider offering a smart    is partitioned into attributes X nt j,F , which pertain to the fixed parameters &#945;, as well as into attributes X nt j,R , which pertain to the individual-specific parameters &#946; n , such that X nt j = X nt j,F X nt j,R . For simplicity, we assume that the representative utility is linear-in-parameters, i.e.</p><p>The distribution of tastes &#946; 1:N is assumed to be multivariate normal, i.e. &#946; n &#8764; (&#950;, &#8486;) for n = 1, . . . , N , where &#950; is a mean vector and &#8486; is a covariance matrix. In a fully Bayesian setup, the invariant (across individuals) parameters &#945;, &#950;, &#8486; are also considered to be random parameters and are thus given priors. We use normal priors for the fixed parameters &#945; and for the mean vector &#950;.</p><p>Following <ref type="bibr">Tan (2017)</ref> and <ref type="bibr">Akinc and Vandebroek (2018)</ref>, we employ Huang's half-t prior <ref type="bibr">(Huang and Wand, 2013)</ref> for covariance matrix &#8486;, as this prior specification exhibits superior noninformativity properties compared to other prior specifications for covariance matrices <ref type="bibr">(Huang and Wand, 2013;</ref><ref type="bibr">Akinc and Vandebroek, 2018)</ref>. In particular, <ref type="bibr">(Akinc and Vandebroek, 2018)</ref> show that Huang's half-t prior outperforms the inverse Wishart prior, which is often employed in fully Bayesian specifications of MMNL models (e.g. <ref type="bibr">Train, 2009)</ref>, in terms of parameter recovery. The generative process of the fully Bayesian MMNL model is:</p><p>where ( <ref type="formula">4</ref>) and ( <ref type="formula">5</ref>) induce Huang's half-t prior <ref type="bibr">(Huang and Wand, 2013)</ref>. {&#955; 0 , &#926; 0 , &#181; 0 , &#931; 0 , &#957;, A 1:K } are hyper-parameters, and &#952; = {&#945;, &#950;, &#8486;, a, &#946; 1:N } is a collection of model parameters that need posterior distribution inference.</p><p>Stochastic variational inference aims at finding a variational distribution q(&#952; ) over the unknown parameters that is close -in the sense of the Kullback-Leibler (KL) divergence <ref type="bibr">(Kullback and Leibler, 1951)</ref> -to the posterior distribution of interest P(&#952; |y). Thus, VB minimizes the KL divergence q * (&#952; ) = arg min q {KL (q(&#952; )||P(&#952; |y))}. Using the evidence lower bound (ELBO) q {ln P(y, &#952; )}q {ln q(&#952; )}, the estimation problem is recast as maximizing the ELBO. For choosing the functional form of the variational distribution q(&#952; ), the mean-field family of distributions (e.g. <ref type="bibr">Jordan et al., 1999)</ref> to factorize q(&#952; ) can be exploited. The main result is that the ELBO can be maximized via a simple iterative coordinate ascent algorithm <ref type="bibr">(Bishop, 2006)</ref>, in which the variational factors are updated one at a time conditional on the current estimates of the other variational factors. Because the ELBO is convex with respect to each of the variational factors, the ELBO is thus guaranteed to converge to a local optimum <ref type="bibr">(Boyd and Vandenberghe, 2004)</ref> and is expected to produce consistent estimates <ref type="bibr">(Wang and Blei, 2018)</ref>.</p><p>For the parameters of a mixed logit model {&#945;, &#950;, &#8486;, a 1:K , &#946; 1:N } the variational distribution can be factorized as (see <ref type="bibr">Bansal et al., 2020)</ref>:</p><p>The optimal densities of the variational factors are given by q * (&#952; i ) &#8733; exp -&#952; i {ln P(y, &#952; )}. However, we found that whereas q * (&#950;|&#181; &#950; , &#931; &#950; ), q * (&#8486;|w, &#920;) and q * (a k |c, d k ) are common probability dis-tributions, both q * (&#945;) and q * (&#946; n ) are not members of recognizable families of distributions (due to the fact the conditional logit kernel does not have a general conjugate prior, <ref type="bibr">Bansal et al. (2020)</ref>)</p><p>and their updates require special treatment. The Delta (&#8710;) method (e.g. <ref type="bibr">Bickel and Doksum, 2015)</ref> provides a simulation-based approximation based on a second-order Taylor series expansion.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4">Results</head><p>Logit-type choice models were used to derive estimates of customers' maximum willingness to pay (WTP) for the features of experimental smart EV charging bundles. These WTP metrics reflect monetary valuation coming from the stated choices and the revealed preference mapping. Table <ref type="table">3</ref>: Bayes estimates of marginal utilities and nuisance parameters</p><p>Whereas the standard MCMC sampler was used for the conditional logit estimates, for the mixed logit model with individual-specific and invariant parameters the VB-NVMP-&#8710; estimator introduced in section 3 was used. The posterior means across both specifications are as expected: on average customers prefer a lower annual fee, fewer hours of control yielded to the electric utility for it to decide when to charge the electric vehicle, larger emission savings, and a higher discount. For those attributes that exhibited evidence of unobserved preference heterogeneity, the mixed logit results include estimates of standard deviations and covariance of the individidual-specific parameters. In addition to posterior means for Bayesian point estimation, posterior standard deviations are also included as a measure of uncertainty. To give an idea of model fit, the loglikelihood function evaluated at the posterior means is also displayed.</p><p>From the conditional logit estimates, assuming a homogeneous sample, we derived a negative willingness to pay of $2.41 (in the annual fee of the program) for each hour increase in the timeframe for which the customer is giving up control of charging of their EV (cf. <ref type="bibr">Richter and Pollitt, 2018)</ref>.</p><p>This negative estimate can be seen as an expected rebate in the annual fee that the customer accepts in exchange for their willingness to delay charging. As supported from current charging patterns, most EV owners leave their cars plugged in at home overnight. For optimal flexibility of smart charging, electric utilities would like to control when charging takes place over that whole period (while respecting the stated charging target). However, the negative estimate of the valuation of hours of control yielded to the electric utility means that customers are less likely to enroll in a smart EV charging program with an extensive period of time where customers are expected to give up control.</p><p>The mixed logit estimates provide evidence that customers exhibit heterogeneity in their response to delay charging. Whereas on average accepting a delay in when EV charging takes place (as measured by an additional hour in the time window where the electric utility controls charging) is associated with an expected rebate in the annual fee (of $2.64, which is close to the conditional logit estimate), other features exhibit a positive valuation. For example, each percent increase in the discount offered as incentive to join the smart EV charging program is valued on average by an incremental $4.21 in associated annual fee. Each marginal pound in emission savings is also positively valued. As previous studies in the area have indicated, inhabitants of Tompkins County seem to be more environmentally conscious. In this study, this fact is represented by a higher valuation of emission savings: a marginal improvement in emission savings from the use of cleaner electricity production is reflected by a willingness to pay additional $2.27 in the annual fee by customers in Tompkins County (cf.the average $1.12 for customers outside Tompkins County). To give an idea of the magnitude of the environmental benefits, an upper bound for emission savings comes from shifting 100kWh for fully charging the largest EV battery from on-peak to off-peak times. For upstate NY, this switching results in roughly 73 pounds of avoided CO 2 e per session. Thus, a key outcome from the mean estimates is that whereas allowing the utility to delay charging is negatively perceived, the negative effect can actually be offset by the incentive (discount) and environmental benefits.</p><p>Although, most of the posterior distributions of the individual-specific willingness to pay measures are consistent with the signs of the mean estimates, the quantiles reported in Table <ref type="table">4</ref> show that there is substantial heterogeneity. Figure <ref type="figure">3</ref> shows, for instance, the conditional posterior distribution of the mean willingness to give up an additional hour in controlling EV charging.  Whereas the most of the individuals in the sample exhibit a negative willingness to pay (expecting a compensation or rebate in the annual fee, as discussed above), some individuals have a positive valuation that is independent of the benefits (discount and emission savings) that are associated with delaying charging. </p></div><note xmlns="http://www.tei-c.org/ns/1.0" place="foot" n="1" xml:id="foot_0"><p>Across upstate New York, within the footprint of the local electricity utility.</p></note>
		</body>
		</text>
</TEI>
