Quantifying Behavioral Impacts on Electric Vehicle Efficiency

The impact of human behavior on vehicle efficiency has been vastly explored for internal combustion engine (ICE) vehicles. However, human behavioral impacts on vehicle efficiency have not yet transitioned to include battery electric vehicles (BEVs). Understanding the impact of human behavior that achieves BEV efficiency is essential globally, as BEVs begin to retain a significant portion of the automotive market share. BEV sales trends in the US have seen consistent growth since 2010, amounting to over 200,000 units sold by 2015. Globally, the total amount of BEVs and plug-in hybrid electric vehicles (PHEVs) is expected to be 40-70 million by 2025. In light of the growth estimates, defining behavior that induces efficient energy consumption when driving BEVs is essential as these vehicles have a traveling distance constrained to 60-120 miles and can require 1-8 hours to attain a fully charged battery at commercial charging stations. With firm traveling distances and long charging times, defining human behavioral impacts on BEV efficiency will allow drivers to get the most range out of their vehicle. In order to develop categories of BEV drivers in terms of efficiency, an empirical experiment was conducted to determine if clustering drivers on their energy consumption profiles invokes significant categories. The driving attributes that defined the clusters were extracted to compare whether or not efficient BEV driving is similar to eco-driving in ICE vehicles. Furthermore, BEV drivers can suffer from anxiety that stems from limited traveling distance, a phenomenon known as range anxiety. However, there exist other sources of anxiety-related human driving behavior, three of which can be measured using the driving behavior survey (DBS). The three anxiety measures from the DBS were contrasted against the BEV efficiency clusters found from this research, to determine if the anxiety factors defined by the DBS were responsible for efficient BEV driving. The results from this research found two significantly different clusters of BEV driving efficiency, which were defined as efficient and inefficient BEV driving. In comparison to eco-driving in ICE vehicles, both aggressive speed and acceleration were found to be contributing factors to BEV efficiency. The results from the DBS proved that anxiety was not a contributing factor to BEV efficiency, as both clusters had similar answers. The information accumulated through this research can be used to guide new BEV drivers to adopt sustainable driving behaviors, which can help maximize their traveling distance on a single charge. Behavioral contributions to efficiency stemmed mostly from reduction of traveling speed; however, consumption based on elevation and road class selection were also quantified. Drivers can use this information to plan their driving routes to minimize energy usage. Modeling techniques that assume a single rate of energy consumption for the population should include behavioral rates defined by this research. Defining behavioral classes of BEV efficiency is essential as BEV sales are on the rise and drivers and manufactures can both use this information to improve efficiency of these vehicles.

as BEVs begin to retain a significant portion of the automotive market share. BEV sales trends in the US have seen consistent growth since 2010, amounting to over 200,000 units sold by 2015. Globally, the total amount of BEVs and plug-in hybrid electric vehicles (PHEVs) is expected to be 40-70 million by 2025. In light of the growth estimates, defining behavior that induces efficient energy consumption when driving BEVs is essential as these vehicles have a traveling distance constrained to 60-120 miles and can require 1-8 hours to attain a fully charged battery at commercial charging stations.
With firm traveling distances and long charging times, defining human behavioral impacts on BEV efficiency will allow drivers to get the most range out of their vehicle. In order to develop categories of BEV drivers in terms of efficiency, an empirical experiment was conducted to determine if clustering drivers on their energy consumption profiles invokes significant categories. The driving attributes that defined the clusters were extracted to compare whether or not efficient BEV driving is similar to eco-driving in ICE vehicles. Furthermore, BEV drivers can suffer from anxiety that stems from limited traveling distance, a phenomenon known as range anxiety. However, there exist other sources of anxiety-related human driving behavior, three of which can be measured using the driving behavior survey (DBS). The three anxiety measures from the DBS were contrasted against the BEV efficiency clusters found from this research, to determine if the anxiety factors defined by the DBS were responsible for efficient BEV driving.
The results from this research found two significantly different clusters of BEV driving efficiency, which were defined as efficient and inefficient BEV driving. In comparison to eco-driving in ICE vehicles, both aggressive speed and acceleration were found to be contributing factors to BEV efficiency. The results from the DBS proved that anxiety was not a contributing factor to BEV efficiency, as both clusters had similar answers.
The information accumulated through this research can be used to guide new BEV drivers to adopt sustainable driving behaviors, which can help maximize their traveling distance on a single charge. Behavioral contributions to efficiency stemmed mostly from reduction of traveling speed; however, consumption based on elevation and road class selection were also quantified. Drivers can use this information to plan their driving routes to minimize energy usage. Modeling techniques that assume a single rate of energy consumption for the population should include behavioral rates defined by this research. Defining behavioral classes of BEV efficiency is essential as BEV sales are on the rise and drivers and manufactures can both use this information to improve efficiency of these vehicles.

ACKNOWLEDGMENTS
First and foremost, I would like to thank my major professor Dr. Gretchen A. Macht for her invaluable guidance in completing this research. Without her unrelenting commitment to make this research both feasible and valuable, the work completed would not have been possible. It has been a privilege to have worked with her over the past year, and those who choose to work with her in the future will quickly learn the passion she has for helping students succeed. I look forward to hearing about her continued success as a professor at the University of Rhode Island.

Background
Human behavior is naturally complex, as behavior is not tangible matter that can be simply measured, but rather a system that changes dynamically in different environments [1]. Beyond studying behavior subjectively, when humans become intertwined with auxiliary systems, complexity grows as the combination of systems forms a hierarchy [2].
Quantifying this hierarchical system with respect to transportation networks is known as driving behavior. Driving behavior is the effect of the interactions between the human, the vehicle, and the roadway system, where identical causes produce variable effects from driver to driver [3].
The field of driving behavior is not new research. As roadway networks became increasingly dense, due to the non-stop growth of registered vehicles [4], the focus of driving behavior research was to define traits that evoke risky behavior which results in traffic accidents and fatalities [5,6,7]. In conjunction with increasing vehicles on the roads, problems arose with a growing amount of pollution from internal combustion engine (ICE) vehicles.
ICE vehicles are responsible for 45% of harmful pollutants emitted every year in the US alone [8]. With respect to human behavior, eco-driving methods were defined as ways to improve ICE efficiency, through the reduction of aggressive driving [9]. The majority of eco-driving protocols call for controlled rates of speed and acceleration; however, further maintenance and comfort settings also inhibit ICE vehicle efficiency [10].
However, as the world's fossil fuel supply lingers [11], the cost to find, extract, and refine oil will continue to grow as supply dwindles [2]. To circumvent a limited supply of energy, alternative fuel vehicles began hitting the market, with increasing sales from year to year.  [4] motor and must be charged at an electronic charging station in order to regain energy for travel [12]. A PHEV is also powered by an electric motor, but has an additional gasoline powered engine used to charge the vehicle's battery pack for extended range [13]. Since a BEV is constrained to electricity as its only source of fuel, traveling long distances becomes difficult, as range is generally limited to 60-120 miles [14]. Once a majority of energy is used for travel, BEVs must be recharged. Recharging the vehicle becomes challenging, as charging times range from 1-8 hours to fully recharge the battery at commercial charging stations [14,15]. Long charging times paired with limited range means BEV drivers have to construct precise driving schedules.
In addition to planning charging events, a BEV driver also needs to consider how much energy they will personally consume. A Mitsubishi i-MiEV has an estimated range of 90 miles [16], but studies with subjects driving the same route had variable usage of energy consumption, ranging from 11% to 15.5% of the total battery capacity [17]. While the variation in energy consumption profiles between drivers could have diverged due to roadway congestion [10,12,18], variables such as traveling speed are known to affect the energy consumption of a BEV [16]. While aggressive driving could have been assumed to be the reason behind inefficient BEV driving, further proof was necessary to support this hypothesis, as BEVs and ICE vehicles are independent systems.

Research Goals
The contribution of this research will serve to quantify the different categories of BEV efficiency, in conjunction with individualistic driving behavior. To complete this analysis, an experiment was set up to collect data on thirty drivers over the course of a 26-mile route.
The route chosen for the experiment was held constant for each driver. Over the course of this route, data was collected via electronic sensors to study BEV driving behavior.
One of the key difficulties for drivers considering switching from an ICE vehicle to BEV is limited range [19]. This notion is known as "range anxiety" which is a measure of anxiety traits that are heightened when the maximum distance a vehicle can travel becomes further constrained, where drivers can be trapped in a vehicle with no battery power and no way to recharge the battery [20]. In an effort to further quantify the effect anxiety has on driving behavior, participants in the experiment took the Driving Behavior Survey (DBS).
The DBS is a questionnaire that has been used to find three levels of anxiety-based driving behavior [21]. If anxiety-based driving behavior was found to be a contributing factor to BEV efficiency, range anxiety could be further detrimental to the efficiency of a BEV driver [22].
Therefore, the following specific research questions were addressed: • Do drivers consume different amounts of energy when driving the same path?
• Does grouping BEV drivers based on their energy consumption profile aid in understanding why some BEV drivers use more energy than others?
• Is anxiety-based driving responsible for inefficient BEV driving?
• Since eco-driving, with respect to ICE vehicles, is assumed to stem from aggressive displays of speed and acceleration, is efficient BEV driving behavior similar to ecodriving?
To answer these specific research questions, this thesis was broken up into following chapters. Chapter 2 provides an outline of the literature encompassing driving behavior, in terms of how behavior is defined, analyzed, and interpreted. An in-depth discussion of the factors essential to eco-driving was explored. Then, BEV efficient driving styles were compared and contrasted to those essential to eco-driving.
Chapter 3 describes the methodologies utilized in this research. Rationale for using the experimental design, along with the process of selecting and recruiting drivers, was defined.
Networking into the vehicle's controller area network (CAN) was used to extract measurements from the vehicle's battery pack [23]. A separate global positioning satellite (GPS) was used to collect location and velocity information. Since smartphone accelerometers have been successfully used to obtain three-dimensional acceleration data [24], an iPhone 6s was used to capture this information. The last section of Chapter 3 covers the methods used to cluster drivers based on their energy consumption profiles, and how the DBS was used to test if anxiety is affecting BEV efficiency within the clusters found.  [3]. When studying driving behavior, research strives to quantify these situations and extract patterns that exist within a segmented population. There exist three key avenues for segmenting drivers with respect to behavior. The three methods are population demographics, psychological traits, and pattern recognition from electronic sensors. However, when it comes to studying human behavior our initial perceptions of a population's performance may be distinctive, but the latent construct of behavioral patterns that define an individual can only be unearthed through the science of behavior [1].
In the process of developing a framework to organize humans with respect to their driving behavior, many initial hypotheses focused on traditional population demographics.
These categories include age, sex, and driving experience [25,26,27], and are typically used to subsegment the entire population into notions of driver categories. Sometimes these population divisions produce significant results. Age and sex show significantly different driving behavior when measuring the gap-acceptance in making left turns [27], and through questionnaires focused on traffic violations and accident risk assessment [28]. However, these divisions are not always clear indicators of driving behavior classes. This can be seen in an experiment to quantify braking behavior at intervals of driving experience, which yielded insignificant results [25]. Success in applying various population demographics in driving behavioral studies exemplifies how a predetermined view of behavior can be used to quantify driving categories. However, the predetermined intuition of these behavioral patterns falls short in some avenues of analysis because human behavior is complex [1] and cannot always be defined by physical traits.
To overcome the complexity that exists in the driving behavior of humans, research has also expanded to incorporate psychological measures. Background knowledge on these traits is usually generated via questionnaires that seek to statistically group drivers based on a group's response pattern [21,29]. Other research has used data from vehicles embedded with electronic sensors to cluster patterns of driving data into known psychological traits [30]. Electronic sensor data is also used for general clustering of driving behavior, by clustering drivers on comparable patterns [5,31].
Exploration of how behavioral patterns vary with respect to driving behavior is further used to understand their impact on roadway systems. Overall, behavioral research in this domain is primarily focused on traffic safety [5,6,7]. Even though traffic safety is a fundamental component of driving behavior research, engineers today are being tasked with alternative avenues of design and analysis by employing green engineering practices.
Green engineering involves quantifying the risk of pollutants through product use and manufacture, in an effort to minimize excessive use of resources [32]. Beyond the design of vehicles, green engineering can be used to reduce ICE vehicle emissions production through analysis of driving behavior. The efficiency of an ICE vehicle correlates with the behavior of different drivers [33]. The application of green engineering to ICE vehicles is known as eco-driving. Eco-driving, also known as ICE efficiency, is affected by aggressive driving behaviors, most commonly noted as aggressive speed and acceleration [34]. There are, however, more factors that are detrimental to ICE efficiency, and understanding those factors will help illustrate how behavioral driving affects efficiency.

Behavioral Impact on ICE Vehicles
It has been estimated that ICE vehicles are responsible for 45% of pollutants emitted in the US [8]. As of 2015 there were more than 263 million registered vehicles on U.S. roads, where alternative vehicles (BEVs, hybrids, and PHEV) accounted for approximately 1.5% of those vehicles [4]. From a green engineering perspective, determining factors to curb emissions from the use of ICE vehicles is the first logical step in developing a more sustainable society. Through the lens of behavioral driving, curbing emissions of an ICE vehicle can be done through eco-driving practices. Eco-driving is defined as the reduction of aggressive driving behaviors, which leads to increased fuel economy [9].
One way to achieve efficient driving of ICE vehicles is to stabilize gear shifting behavior. In a study that simulated gear shifting relative to speed, it was found that an aggressive style of gear shifting can increase fuel consumption and CO 2 production by up to 30% [35]. To apply non-aggressive gear shifting, vehicle operators would need to up shift between the revolutions-per-second (RPM) range of 2000-2500 [36]. When an ICE vehicle is up shifted sooner, wasted energy used to propel vehicles on lower gears at faster speeds can be avoided.
A second method to exercise eco-driving principles is to control one's rate of acceleration and deceleration [36]. Ideally, drivers would seek to control braking and acceleration behavior when driving in variable traffic congestion levels [37,38]. ICE vehicles make use of both automatic and manual transmission, so engine braking can be used to slow down vehicles when approaching a stop signal or entering roadways with traffic congestion. The engine braking method utilizes smoother deceleration, which can aid in reducing speed appropriately, without going below the optimal speed, which would then require more unnecessary acceleration [18]. Reducing the rate of acceleration boils down to maintaining consistent speeds, which ultimately increases fuel economy of an ICE vehicle [37]. Choice of traveling speed, however, also affects efficiency. When traveling speed is compared against fuel efficiency, a negative parabolic trend occurs [10]. For example, a V6 2007 Honda Civic exhibits minimal efficiency at speeds of 30 and 90 miles per hour (MPH) and optimal efficiency around 61 MPH [39]. What this means for eco-driving efficiency is that road type selection is important when seeking ICE efficiency [10,37].
Other known factors affecting ICE efficiency include excessive idling, which can account for a quarter to a half gallon of fuel per hour [40], and upkeep in regular maintenance, by tuning the engine, keeping appropriate tire pressure, and selection of engine oil [10].
Additionally, the mass of the vehicle can impact overall efficiency, because of the energy requirements necessary to increase momentum [10]. Lastly, usage of cooling by means of heating, ventilation, and air conditioning (HVAC) systems can reduce mileage by 5-25% [41]. However, heating is not affected in the same manner because ICE vehicles heat is generated from waste engine heat, where no additional energy is used to heat the vehicle's cab.
While all of the factors affecting ICE efficiency can be determined through mathematical modeling or mechanical systems simulation, understanding the human behavioral interaction with ICE vehicles for efficiency can only be achieved through experimentation.
Since mass is an important aspect of ICE efficiency, buses are affected not only by their size, but also by their capacity of travelers [42]. Buses can be more efficient in a sense that per rider, their emissions are lower than driving individual vehicles because buses are a form of carpooling [10]. However, other factors still play a role in determining a bus driver's efficiency. From an experiment tracking three bus drivers, on five different bus routes, it was found that each driver used a different amount of fuel on the same route.
There also existed consistency among the drivers, where one driver always used the least amount of fuel, another used the most, and the last driver was always in between the two [43].
However, utilizing buses as a mode of transportation is not nearly as common in the US as it is in European countries [44]. As it pertains to studying driving behavior with respect to eco-driving of passenger vehicles, most research is focused on testing whether or not individuals' driving behavior can be altered to improve efficiency [36,37,45]. The reason that individualistic patterns of eco-driving may not have been exhaustively researched is because aggressive driving behavior was stated to be the reason for inefficiency [9]. The gap in research could also stem from the US eliminating eco-driving programs which informed drivers on how to improve their efficiency [9]. Another issue with eco-driving is permanent adoption of the habits, when there exist minimal negative effects from drivers fading back to their standard habits. One experiment found seven out of eight subjects adopted eco-driving habits over the course of 6 months after taking an eco-driving course [36]. However, since adoption of these habits can be affected by driving situations and personal motivations [47], without a more complete sample size adoption of eco-driving for the long term has not completely been researched. It may also be challenging mentally to adopt eco-driving habits because ICE vehicles have been seeing an upward trend in efficiency. The Environmental Protection Agency (EPA) tracks the fuel economy and the rate of emissions from all vehicles in the US [46]. The EPA's data, seen in Figure 2.1a, shows an uptrend in the average fuel economy of cars (green) and trucks (blue). This increase in fuel efficiency also correlates with a downtrend in the rate of of cars and trucks has increased by 29% and 27% respectively. The increasing trend of efficiency is due to governmental policies that set standards for fuel economy and production of emissions [48]. As of 2016, the average fuel economy of cars was around 29 MPG. By the year 2020, the National Highway Traffic Safety Administration has set standards for cars to achieve 35 MPG. Individual vehicle manufacturers are responsible for meeting set standards; otherwise the manufacturer is fined for failing to meet expected efficiency [48].
With an upward trend in efficiency, adoption of eco-driving behaviors indefinitely will be even more challenging for society as ICE vehicles become more efficient on the man-  [49,50]. While that number may be difficult to reach, the rate of BEV sales over the last five years continues to increase in the US, from 9,750 units in 2011 to 71,044 units in 2015 (see Figure 2.2). To prepare for this rising green technology, it will be essential to understand driving behavior's impact on BEV efficiency before these vehicles retain a larger market share, as compared to studying behavior of ICE vehicles which is being done retrospectively.

Behavioral Impacts on BEVs
To apply behavioral techniques to BEVs, as has been done with ICE vehicles through eco-driving, will first require knowledge of how BEV efficiency is influenced. While both vehicles look similar, their internal systems are substantially different. The motor of an ICE vehicle has hundreds of moving parts, as compared to a BEV which only has an encompassing motor [51]. A rough schematic of the essential components for each vehicle is depicted in completely different, making them unique systems [2]. In studying ICE efficiency, a focus on the variables that affect the rate of fossil fuel consumption is monitored [36,37]. For a BEV, a focus on the variables that affect the rate of sate of charge (SOC) depletion, must be studied to determine the impact of human behavior. The first variable affecting the rate of energy consumption, and one of the most important, is vehicle speed. An experiment was conducted using a Mitsubishi i-MiEV and a Nissan® LEAF®. A single driver, operating each of the cars at speeds of 37.2, 43.5, 49.7, 55.9, 62.1, and 68.3 MPH, found an exponential relationship between speed and rate of energy consumption [16]. This curve is much different when compared to the speed vs.
fuel consumption curve of an ICE vehicle, which follows a negative parabolic trend [10].
In terms of efficiency, an ICE vehicle is efficient around 45 MPH, and a BEV is efficient around 15 MPH [12]. Since efficiency follows two completely different trends with respect to speed for ICE vehicles and BEVs, road type selection between the two is completely different.
Energy savings via road type selection for BEVs was found by collecting data over an extended period of time from a professor's daily commute. Initially there were four possible ways to get between the two places (home and work). The three main routes used were comprised of back roads and intercity roads, while the fourth route used a highway for travel. The time savings from using the highway were minimal compared to the other three routes; however, the other three routes saved about 1 kWh of battery capacity over 100 miles [53]. Utilization of alternative routes can be thought of as leverage points, where the length of time may not be a significant factor, but the rate of discharge of the BEV can be beneficial to the driver [54].
Road grade is another important factor in route selection. BEVs have the ability to regenerate energy while braking. The regenerative braking system allows vehicles to harness the power lost while braking, by reversing the direction the motor turns [55]. It is estimated that in urban areas, the recuperation from braking could increase the range of a BEV by 15-20% [12,55]. The effect of power used by a BEV versus the traveling grade was studied, and showed that a positive increase in grade always has a significant amount of power consumption, while a decreasing grade results in minimal energy consumption or energy collection from the regenerative brakes [53]. In contrast, ICE vehicles' efficiency suffers from elevation changes, because they cannot recover energy on decreasing road grades. In fact, a study found that an ICE vehicle traveling the same distance on a flat route and a route with variable elevation had fuel savings of 15-20% on the flat route [56].
Regenerative braking can also increase efficiency in other scenarios, such as driving in variable traffic congestion, because of the deceleration. A study was conducted on various levels of traffic congestion, where measurements were recorded during free flow traffic, mild congestion, moderate congestion, and high congestion. When a freeway switches from free flow to mild/moderate congestion, BEVs take advantage of slower speeds and regenerative braking though deceleration for improved efficiency [12]. When compared to an ICE vehicle, which is not efficient at slow speeds, the rate of fuel consumption increases [10]. Another problem in high traffic congestion is that if traffic comes to a full stop, ICE efficiency is reduced because of excessive idling [40]. A BEV's motor, on the other hand, does not turn without pressing the accelerator; thus, no energy is consumed from the motor when idling.
Another factor affecting a BEV's rate of energy consumption is the auxiliary power used by systems such as steering, radio, HVAC, and other onboard electronics that derive power directly from the BEVs battery. Since in a typical ICE vehicle these amenities are powered via an alternator, their effect on fuel consumption is not as detrimental as it is to BEV drivers who have to account for the charge they will use on a trip. For a Nissan® LEAF®, it has been estimated that the power needed to support the accessories is 0.2 kW, while the power needed to sustain the HVAC system is around 6 kW [57]. The HVAC system affects efficiency of BEVs at both ends of the extreme temperature spectrum [57], where the trend of efficiency is parabolic. When using the HVAC system, a BEV is extremely inefficient at temperatures below 0 • F and above 100 • F, and is most efficient around temperatures of 50 − 70 • F [58]. Overall, HVAC systems have been found to reduce range by up to 40% [59].
For the most part, the variables that affect ICE efficiency are different from those that alter a BEV's efficiency. This is because ICE vehicles and BEVs interact with transportation networks differently. For ICE vehicles, it has been suggested that aggressive driving behavior is directly linked to ICE efficiency [9]. While this assumption may hold true for BEVs in the case of the exponential use of energy at higher speeds [16], aggressive driving is also affected by rapid rates of acceleration [34]. All BEVs take advantage of regenerative braking and recover energy during deceleration [53], making the overall assumption that aggressive driving is the sole representation of BEV efficiency challenging to prove without experimentation.
However, what is known about BEVs with respect to driving behavior is that consumption of energy differs between drivers. An experiment having ten drivers drive a BEV on the same route resulted in variable usage of energy among the participants [17]. Another study's findings showed variation in the rate of SOC consumption among a group of 25 drivers, where the least efficient driver consumed 2.8% SOC/mile and the most efficient driver used 1.5% SOC/mile [60]. Since it can be interpreted that behavior has an effect on the rate at which energy is depleted, it has been stated that future works should study the behavioral patterns that lead to excessive energy consumption [17]. Research defining behavioral impacts on BEV efficiency can have a greater impact on society as compared to studying eco-driving with respect to ICE vehicles, because limited traveling range may force drivers to adopt efficient driving styles for the long term.
One of the most common drawbacks from society adopting BEVs is their limited range [19], as most BEVs are restricted to a traveling distance of 60-120 miles [14]. However, most BEVs on the market have been found to meet a large percentage of individuals' travel needs [61,62]. A study supplying BEVs to 40 participants for daily driving found that 94% of participants had enough range with the BEV to meet their needs [63]. However, issues with a BEVs limited range are still present. This phenomena is known as "range anxiety," which is defined as"the fear of becoming stranded with a discharged battery in a limited range vehicle" [20]. For BEVs, range anxiety can be mitigated through optimally designing a transportation network with an appropriate amount of charging stations [64].
However, since charging times can range from one to eight hours at commercial charging stations [14,15], and with optimal networks still in the research phase [64], understanding the behavioral component can help BEV drivers learn to get the most out of their vehicles' range to reach destinations and charging stations.
Since range anxiety can be overcome through BEV driving experience [62], and since studies point to high interest in BEV usage [65], studying the impacts of BEV behavioral driving can help plan for a transition to a BEV future. There is a strong possibility that there will be between 9-20 million electric vehicles (EVs) by 2020 and between 40-70 million EVs in the world by 2025 [66]. By quantifying the impact of BEV driving behaviors, current BEV drivers can learn to practice efficient driving behavior and new BEV drivers can be educated to help make a seamless transition into the future.

Methodologies
In order to appropriately define classes of BEV drivers with respect to efficiency, various methods were deployed to collect and analyze data. This chapter begins with the methods used to develop an experiment where subjects drove a BEV for the bulk of data collection. Thens the systems used to electronically collect data from driving and use of questionnaires were discussed. Finally, the methods chosen to analyze the data will were covered.

Methods of Experimentation
This section will cover all of the methods used to design and run an experiment to capture data relevant to BEV driving behavior. For this experiment, a sample of subjects was recruited to drive a BEV on a predetermined route.  [49,50], the impact from studying the driving behaviors of inexperienced BEV drivers was of greater benefit to society, as there is a limited pool of experienced BEV drivers and their habits will need to be quantified to improve their BEV driving efficiency. In general, route selection for studying driving behavior has been held constant in previous research, as seen in the constants column of Table 3.2. For this research, the route driven by the subjects was also held constant. The importance of holding the route con-stant was necessary for this research, because assessing total energy consumed over the route was an essential step in defining behavioral patterns. The only way to make these comparisons was to hold route characteristics that affect the rate of SOC depletion, such as elevation changes [12] and road types [53], constant. However, the experimental route did include those characteristics, as it was necessary to interpret how these factors affect BEV driving behavior.
Since road type selection is a component of BEV efficiency [53], the route adopted for experimentation covered various road classifications. To make distinctions between one road and another, the road classification practices presented by OpenStreetMaps (OSM) were used. OSM is an open source, community-built database of roadway information [69]. The descriptions of the different road classes are available in Table 3.3. The road classes from OSM were all plotted in southern Washington County, RI, as seen in Figure   3.1a. The chosen route, depicted in Figure 3.1b, was eventually selected as the route of choice. The route covers every OSM road class, except for Motorways. In the region where the experiment took place, that road class did not exist, as is not seen by the missing color red in Figure 3.1a. In the experimental route, participants started at the University and  The average distance on a given road class was 5.29 miles, with a standard deviation of 0.91 miles. Two road types that were split between two segments of the route were Minor Arterial (orange) and Minor Collector (yellow). On the Minor Collector road class, the total distance for the first portion was 1.68 miles and the second portion was 4.31 miles long.
For road class Minor Collector, the first portion was 2.25 miles long and the second portion was 1.82 miles long. Even with those splits, there was still a large enough distance on each split to compare to the intervals of change in percent battery capacity, as this happened, on average, every half mile.
While maintaining road types was integral, the change in elevation along each road type was also considered. Since the area in southern Washington County had a high variation in elevation, the design of the experimental route was planned on roadways that displayed variable elevation changes. Beyond the experimental route, other conditions were held constant for each driver in the experiment. The first constant was weather. The experiment was not held on rainy days, by pre-scheduling experimental drives on days where the weather forecast did not show rain. In the event that a driver was scheduled for a morning with rain, they were rescheduled for another day, or later in the day if the forecast was expected to be sunny for more than four hours. The last variable held constant in the experimental drive was the brake recuperation mode. The e-Golf has the possibility to recuperate energy at four different levels, which are outlined in Table 3.4. While braking in a BEV always recuperates some amount of energy, these modes apply a degree of braking power when the driver's foot is removed from the accelerator. Modes D1-B in Table 3.4 range from light braking to very high braking, meaning that once the driver removes their foot from the accelerator, the brakes will be applied to the degree of the recuperation mode. Mode N allows drivers to experience normal braking, where the brakes are only applied when the brake pedal is pressed [70].  Before subjects began driving, they were instructed to attempt to recreate a scenario that promoted their preferred vehicle setup. This included adjustment of the seat and mirrors, if the subject felt it was necessary. Subjects also had the choice to listen to the radio, or their own music. Since this option was given to everyone, the center console display was always left on, to retain a similar level of energy draw, from those components, from the battery.
Since the participants were first time BEV drivers, they were each given a brief introduction to the car before starting the experimental drive. This included how to visualize their energy consumption, which could be done in two ways. The first was by looking at the SOC gauge from The DBS was used in this research to measure the effect of anxiety on BEV driving behavior. A discussion of the DBS was covered in the last section of this chapter. However, since a portion of this experiment was focused on studying the factors of anxiety presented by the DBS, range anxiety was not measured in the experiment. In an effort to curb this type of anxiety, participants were informed that they had plenty of energy in the BEV to complete the experiment.
At the conclusion of the experiment, two additional control runs were used to capture data for further analysis. For each control run, constants in the experiment remained the same. The first control drove the BEV efficiently, while the second control drove the BEV inefficiently. For efficient driving, a focus on maintaining controlled speeds equivalent to the speed limit and focus on minimal consumption based on the kWh gauge was employed by the efficient control, along with non-aggressive acceleration, as this was defined to be an aspect of eco-driving for ICE vehicles [9]. The inefficient control drove the BEV using aggressive speed and acceleration, while focusing on high consumption from the kWh gauge. Use of additional control drives have been used in similar experimental designs to establish separation between potential behavioral clusters [31].

Methods of Data Collection and Analysis
This section covers the various sources of data extracted during and after the experiment. During the experiment, electronic sensors were set up throughout the vehicle, taking measurements from the vehicle's computer and some external measurement systems that recorded information, such as location and acceleration. A discussion of the questionnaire taken after the experimental drive was also covered in this section.

BEV Driving Data
The most crucial element for the completion of this research was a BEV. The vehicle selected for data collection was a 2015 VW e-Golf. This vehicle was equipped with an 85 kW electric motor, which is powered by a 24.2 kWh lithium-ion battery pack. The manufacturer specified range was around 85 miles and the vehicle had a curb weight of 2.455 tons [71]. The selection process for using this vehicle stemmed from both its competitiveness with other BEVs on the market in terms of driving range, and from the fact that no other literature found had used a VW e-Golf for electric vehicle experimentation. Most research studies used a Mitsubishi i-MiEV, which has an estimated range of 90 miles [16,17] or a Nissan® LEAF® which has an estimated range of 123 miles [16].
The first piece of data necessary for this research was the SOC of the vehicle. SOC is the percentage of battery power remaining in the battery pack before the BEV needs to be recharged. While this information is available from an analogue gauge on the dashboard,   Figure   3.4b. This board is called the PiCAN2 which transformed the Pi into a CAN controller.
Wires were connected from the CAN-High and the CAN-Low terminals on the PiCAN2 board, which were then linked to the terminals on the OBD port on the BEV. These wires could also have been connected to the CAN-High or CAN-Low wires at other junctions on the car, by splicing those wires and pig-tailing them together. The setup for the system can be seen in Figure 3.4b, where the Raspberry Pi 3 is the green board on the bottom, connected to the blue PiCAN2 board on top.
With the Raspberry Pi connected to the vehicle, CAN frames were read directly from the BEVs computer. The frames transmitted via the CAN wires contain segments of messages represented as binary strings. These strings are segmented into groups enabling the BEVs computer to interpret their meaning. Examples of these groups included the start-offrame bits, which let the computer know a new frame is being transmitted; end-of-frame bits that let the computer know that a new frame will begin now; and the data bits, which hold measurements from different sensors on the BEV [73]. Using SocketCAN to read this data enables programmers to split these fields automatically, without having to read raw binary strings to ascertain data from the system.
The output depicted in Figure 3.5 is the output from the Linux terminal when the Can-Dump command was issued [75]. The first column of values in Figure 3.5, highlighted in green, displays the name of the node assigned to the Raspberry Pi. The second column, highlighted in cyan displays the Parameter ID (PID). A PID is a data location identifier.
Every CAN system has a set of unique PIDs, which are the identifiers for data stored in a frame. For example, PID A5 could store both speed and RPM. Every time the system witnesses PID A5, the controller reading that frame will know that the frame stores speed In this case the number of bytes specified in the third column represents how many columns of bytes will follow.
The data shown in Figure 3.5 mean nothing on their own. However, to the vehicle they supply the necessary information to all electronics on board to make the car function. The vehicle's internal database keeps a record of what each CAN frame means, and how it will be decoded by subsystems in the car. Obtaining the SOC data required finding the PID that identifies the data packets containing SOC measurements, by decoding the hexadecimal data packets from the yellow highlighted section in Figure 3.5. There were two possible methods for decoding the frames, known as fuzzing and visual pattern correlation [74].
Fuzzing would require manually creating CAN frames and injecting them into the system, which was possible to do in SocketCAN. If a custom frame was injected into the vehicles CAN, the vehicle would react in some manner. Simple examples of a given reaction would have been the horn honking or the speedometer jumping to a different speed. In this research, a change in the SOC gauge would have been sought after when fuzzing CAN data (see Figure 3.6a). While fuzzing would have been the fastest way to decode the frames, it is also very dangerous as the vehicle may act erratically which could have risked damage to the car or anyone in the car [74].
To circumvent the issues with fuzzing this research employed method two of CAN frame decoding, which was visual pattern inspection. One way to have applied this method was perform an operation on the car, and look for bytes of a given PID that changed. For example, gas pedal and brake pedal percent depression were decoded in this fashion with only the interior power switched on. The pedals were pressed, and the changing bytes were found. The benefit of only having the interior power on was that many other CAN frames were unchanged; thus, they could be filtered out, leaving only a select few frames that changed. For SOC, however, this method would not work because there was no way to rapidly discharge the battery safely without driving. Thus, every CAN frame received did not witness a change in SOC, as can be seen in Figure   3.6b, where the data presented themselves as a stepwise function.
Location of the BEV was a requirement of the data collection. While it was noted that CAN data for SOC does not change with a high degree of frequency, it will be beneficial to sample location information comparable to the speed put out by the CAN, so that other sources of data such as accelerometer and engine load measurements could be accurately Another device used to collect data from the vehicles was an Apple iPhone 6s. Smartphones are equipped with accelerometers because these devices can be used to track screen orientation. Smartphones have become standard in vehicle driving based experiments [24,68,77]. While a smartphone accelerometer's measurements present a greater degree of noise, they produce correlated results to commercial devices [24].
The iPhone application Sensor Play was used to log data. The application allows for a sampling rate of 10Hz, in line with both the CAN sampling rate and the GPS sampling rate. The iPhone was positioned on the cup holder tray, and locked down to prevent movement. Since the iPhone was well oriented, readings did not need to be reoriented [77] as the iPhone was held stationary. The spatial measurements of acceleration taken from the iPhone are illustrated in  Once all of the datasets were fused, there were some further post-processing steps. The first was to calculate the distance between each logged event. This can be done using the Haversine formula from Equations 1 and 2, which equates the distance between two geocoordinates on a sphere. R is the radius of earth, represented in miles, and is equal to 3,959 miles. ϕ and δ are the latitude and longitude coordinate values respectively. For each coordinate value a subscript of 1 represents the previous set of coordinates (look back one row) from the combined dataset. A subscript value of 2 was equivalent to the current coordinate in the combined log file.
With the distance between coordinates calculated, these values were then used to distribute the measurements of SOC backward, from the point in time they were logged to the last change measurement in SOC. This process was necessary because the rate of SOC depleted was studied with respect to road classifications. Since the logging of SOC was found to be a stepwise function, there could have been over-or under-inflated measurements of change in SOC due to carry-over between road classes. To circumvent this issue the first change in SOC for each change in road class could have been stripped from the dataset for each subject. This would have amounted to a significant loss in data. Instead, change in SOC values were backwards distributed based on distance traveled.
The process for this backwards distribution of SOC measurements is illustrated in Figure 3.9. Incremental values were calculated using ∆SOC ∑ Distance , where ∆SOC is the change in SOC, reported by the vehicle, represent as 0.1 kWh in Figure 3.9. Then each block of distance was multiplied by the incremental value. This way carry-over between roads will not be over-or under-inflated.
Once the data was cleaned, analysis of BEV driving behavior transpired. To categorize drivers with respect to their rate of SOC depletion, cluster analysis was used. Cluster  [78]. There exist a few approaches to complete this form of analysis. K-means clustering is by far the most popular; however, it requires sample sizes greater than 200 [33,78,79]. For this reason, hierarchical clustering was used.
Hierarchical clustering is beneficial for sample sizes smaller than 50, as the decision between clusters is done through visual inspection of the separation of cluster on a dendrogram (hierarchy tree diagram) [80]. To conduct hierarchical clustering, a proximity matrix between data points was computed. The values used to compute the matrix were a subject's mean and standard deviation ∆SOC Mile . Since one of the objectives of this research was to group drivers based on their rate of energy consumption, Euclidean distance was used to calculate distance in a two-dimensional plane to measure dissimilarity of clusters. The steps for using hierarchical clustering are as follows [78,81]: 1. Start with all observations (n) belonging to their own singular cluster 2. Combine two clusters that produce the smallest impairment to the objective function; total clusters equal to n -1 3. Repeat step 2 until every cluster is combined into one cluster Step two in the process requires an objective function which is responsible for measuring dissimilarity between clusters. There exist a few possible objective functions, one being single-link, which chains clusters based on the minimum distance between a single point in two clusters [78]. Complete-linkage is similar to single-link, except the maximum distance between two points in distant clusters is used to measure dissimilarity [78]. There is also average-link, which measures the average distance between all points in a cluster [78]. However, for this research the Wards objective was used. This method joins clusters based on the smallest error in the sum of squares (Equation 3).
All data in the results and discussion chapter of this paper were first tested for normality Aggressive acceleration patterns were studied by using the safe driving region within a friction circle [84]. a(x) = 0.446 * a(y) 2 + 2.395 * a(y) − 3.349 (7)

Survey Data
The last portion of data collected from the experiment was a questionnaire. There exist many questionnaires with respect to driving behavior. One is known as the Driving Skills Questionnaire (DSQ). The DSQ measures how an individual perceives their driving skills.
The questionnaire consists of 20 questions, where subjects rate how well they believe they perform a given skill on a scale of 0 (very poor) to 10 (very good) [29].
Another questionnaire, known as the Driving Behavior Questionnaire (DBQ), has subjects rate themselves on how often they perform silly or bad driving behaviors. The questionnaire is composed of 50 questions, developed to cover the categories of slips, lapses, mistakes, unintentional violations and intentional violations. The questionnaire uses a 5point Likert scale, where participants measure how frequently an event happens to them while driving, ranging from 0 (never) to 5 (nearly all the time) [85].
The last questionnaire considered for the research was the DBS. This survey measures anxiety with respect to driving behavior. Subjects taking the survey will report how often they react to an anxious situation, described by each question. The questionnaire also employed a Likert scale; however, the scale ranges from 1 (never) to 7 (very frequently) [21].
The three proposed questionnaires all divulged different approaches to defining driving behavior classifications. The DSQ presented itself as an opportunity to test how perceived driving skills affect energy consumption when driving a BEV. However, while the DSQ can result in four factors, subjects tend to over inflate their driving skills [29]. The DBQ could also have been used as it can measure safe/aggressive driving practices [85]. However, these factors could already be determined via the electronic sensors equipped to the BEV [9,34].
The DBS was eventually selected as the best questionnaire to include in the study. However, the DBS was not selected only as a last resort. The DBS measures driving behavior with respect to anxiety [21], and a major factor inhibiting BEV growth is range anxiety [20,62]. If anxiety is the responsible for BEV efficiency, then drivers can run into situations where two factors of anxiety are overwhelming one another. Anxiety can further impact anxiety, developing a vicious cycle that can be overly detrimental to an individual's health [22].
In developing the DBS, two studies were conducted, creating a 21-question survey used to rate a driver's anxiety [21]. In both studies, three factors were established to model anxiety with respect to driving behavior. The three anxiety factors determined by the DBS are outlined in Table 3.5. To ensure that the factors extracted from the DBS differ from other questionnaires, the four factors from the DSQ [29] were compared to the three factors of the DBS, showing that the factors were significantly different [21].
For completeness, all 36 questions used in developing the DBS were answered by each subject in the experiment (see Appendix C). The 21 questions essential to the DBS were The ability to control anxious behaviors in complex driving situations Anxiety-Based Aggressive/Hostile Behaviors (ANG) Accident related fears and eruption of anger episodes extracted and analyzed. By taking the average of the 7 questions that belong to the three categories of anxiety, anxiety can be measured by a low mean suggesting minimal anxiety, and a high mean suggesting heightened anxiety [21]. The means of all clusters of BEV driving behavior were compared against one another. This approach was applied in similar research, and has also been defined as the method of comparing groups against one another using the DBS [86]. Principal component analysis was used to ensure that responses to the questionnaire, taken in this research, loaded correctly on the factors outlined by the DBS.

CHAPTER 4 Results and Discussion
This chapter provides a comprehensive analysis of the data extracted from the experiment. Once a discussion of the raw data was covered, a cluster analysis was used to generate clusters of similar driving behavior patterns. Lastly, an analysis of how the DBS can be used to understand these clusters was assessed.

Analysis of SOC Consumed
In similar experiments, results had exposed that when a sample of drivers was tasked with driving a BEV on the same route, a dissimilar amount of energy was used among the samples [17]. Since the BEV brand used in this research had not yet been utilized in any experimentation, with the extent of literature discovered, the first portion of the analysis determined if an inconsistent usage of energy existed between the subjects. For a visual perspective, a bar plot of the total consumption used by each subject was generated, which can be seen in Figure 4.1. The x-axis labels each of the subjects in the order they were tested, while the y-axis displays their total energy consumption for the entirety of the route in kWh.
Subject 37 was the control for inefficient BEV driving and subject 38 was the control for efficient BEV driving. The efficient control used 6.41 kWh and the inefficient control consumed 7.26 kWh which suggested that there was a difference between the two BEV driving styles. The lower control however did not under-perform other subjects. The least amount of kWh used on the route was done by subject 29, who consumed 1.21 kWh less than efficient control. The maximal difference existed between subjects 3 and 29, amounting to a difference of 2.42 kWH. If the rate of depletion of SOC could effectively be estimated as Miles kW h , subject 29 would have been able to drive an additional 12.28 miles before consum-ing an equivalent amount of energy to subject 3. For the sake of visualizing this difference, subject 29 could have been able to complete an additional 47.23% of the experimental route before using a similar amount of energy to subject 3. 1 2 3  5  7  8  9  10  11  12  13  14  15  16  17  18  20  21  22  23  25  26  28  29  30  31  32  34  35  36  37  38 Subjects (Excluding Incomplete Data) In similar BEV driving research, it was found that subjects had a mean energy consumption of 0.21 kW h Mile , with a standard deviation of 0.04 kW h Mile [17]. In this research the mean consumption was 0.25 kW h Mile , with a standard deviation of 0.02 kW h Mile . While the mean consumption in this research differed by around 0.04 kW h Mile , the difference most likely stemmed from the experimental route in this experiment being much longer, with more elevation changes. Also, the other experiment focused on routes with many traffic signals which would aid in energy recuperation from braking [12,17].

Road Type Analysis
Since it had determined that road type selection for a BEV is essential for reduced energy consumption [53], a graphical representation of energy consumption by OSM road class was created and can be viewed in Figure 4.2a. The x-axis lists the subjects, while the y-axis is scaled by kW h mile , where each subject had 5 data points, one for each OSM road classification. While Principal Arterials (purple) and Expressways (black) were expected to be highly inefficient due to higher speed limits, Major Collectors (blue) proved to be generally the most inefficient type of road to drive on. The interaction plot in

Elevation Analysis
The inefficiency of road class Major Collector was evidence that change in elevation was a key component of BEV inefficiency [12]. Since there existed a limited number of Major Collector roads in the region, the experiment was only conducted on uphill elevation changes, as seen in    For this reason, analysis of BEV behavioral driving with respect to elevation was eliminated from further analysis, and driver behavior was analyzed instead on a driver's overall rate of energy consumption per mile.

Cluster Analysis
From the analysis of the raw data, there existed evidence that subjects were not using a similar amount of energy while traveling the same route, although some subjects appeared to show similar consumption patterns. This section first covers the process of subdividing drivers into clusters through hierarchal clustering. Next, clusters of drivers were statistically validated to ensure that clusters had unequal distributions of energy consumption.
Lastly, driving parameters such as speed and acceleration were tested to determine if the clusters found had displayed aggressive behaviors, as aggressive speed and acceleration are responsible for ICE inefficiency [9].

Results
To cluster drivers based on their energy consumption profiles, hierarchical clustering was utilized. Hierarchical clustering develops a tree-structured dendrogram where clusters can be assigned by the partitions in the branches. The dendrogram from this experiment can be seen in

Cluster Validation
Since three possible distinctions of BEV driving behavior were found, the strength of cluster's to stand on its own as significant categories of BEV driving behavior was covered. To determine statistical significance between found clusters of BEV efficiency, each clusters distribution for their rate of energy consumption was tested for normality. Both an Anderson Darling and a Shapiro Wilk test were used to test the hypothesis of normality of the cluster's distributions.
Every cluster, in a two and three cluster setup, were found to be not normal. All test statistics from the Anderson Darling and Shapiro Wilk tests were zero, meaning that the assumption of normality was rejected. For further inspection, qqplots were drawn for each cluster in a three cluster model. If data were to be considered normal, data points would have linearly trended around the fitted line in a qqplot. However, this was not the case for the data extracted from this experiment. All of the qqplots, seen in Figure 4.5, are non-linear. The curves from each qqplot suggest that the distributions were skewed to the right. The skewness of the distributions seemed truly representative of the data collected, as there were events such as traffic and variable elevation that aided in BEV efficiency, which altered the rate of energy consumption from the BEV [12]. Normal data obtained from BEV driving would more likely exist in non-complex transportation networks, as there would be no event that would hinder the rate of energy consumed. While normality could have been assumed from the implications of the central limit theorem, the data had been proven to be not normal, and further analysis was built on this fact.
Since normative testing between two samples was not used to find significance between the found clusters, a non-parametric test was used. The test used on these data was a Wilcox test, and the resulting test statistics for each type of BEV driving can be found in Table 4.1.
In a two cluster scenario, both efficient and inefficient BEV driving were found to be significant, as the test statistic from the Wilcox test was zero. In a three cluster scenario, significance testing was computed for both efficient behavior and moderately efficient be-  their test statistic was 0.001. Testing was then computed between moderately efficient behavior and inefficient behavior. Since these two clusters were formed by dropping a level on the hierarchical clustering dendrogram (see Figure 4.4a), significance between these two groups could only result in a three cluster model. The resulting Wilcox test statistic, however, was 0.161, meaning that these two groups were highly similar.
The results from statistical testing deemed a two cluster model accurate for driving behavior, within the sample of participants from this experiment. From Table 4.1, the number of subjects is well balanced for a two cluster setup. In the three cluster setup, there were only five subjects that fit into Inefficient BEV driving, so it could have been possible that with more driving samples, three clusters could be significant. However, there was a still a large number of data samples for all clusters which help aid in the accuracy of these results.  which mimics the demographics of the population, as there were more males enrolled in the experiment than females.
Since the two cluster setup was deemed significant, from this research, BEV drivers can either be categorized as efficient or inefficient. Further analysis of how strong these two clusters of behavioral BEV driving was completed by taking the same parameters they were clustered on, those being mean and standard deviation ∆SOC Mile , and separating their consumption on the different road classifications defined by OSM. These drivers displayed efficient driving behaviors on those road classes. This was to be expected, as drivers who were categorized as inefficient BEV drivers had a much higher standard deviation in energy consumption; thus, they will display a degree of variability in energy consumption when driving a BEV.
One road class that did not hold up well was class Minor Arterial. On the Minor Arterial roads, depicted in Figure 4.6c, there was a large amount of mixing between the regions of inefficient and efficient BEV driving behavior. Both control runs also displayed a very similar pattern of consumption over this road class. Road class Minor Arterial, however, had greater traffic density and many stop signals, which provided ample opportunity to recuperate energy from the regenerative brakes.

Analysis of Aggressive Driving Attributes
Part of the analysis of this research was to determine if eco-driving of ICE vehicles is comparable to efficient driving of BEV vehicles. The last portion of analysis of data extracted from the vehicle was used to compare eco-driving to BEV driving efficiency. Ecodriving in ICE vehicles was defined as the reduction of aggressive speed and acceleration.
With this in mind, acceleration data collected from an iPhone 6s and velocity data captured from the Adifruit® Ultimate Breakout board were analyzed.
With traveling speed being a large contributing factor to inefficient driving of BEVs [16], a distinction was made between the traveling speeds of efficient and inefficient BEV drivers. which can be seen in Figure 4.7. Generally, efficient BEV driving behavior (green) results displayed much slower speeds than inefficient driving behavior (red). There was a slight overlap (brown) between the two classes of BEV driving behavior; however, for the most part it would appear that each class's distribution of speed differs. Surprisingly, there is a large degree of overlap on road class Expressway, which can be seen between time steps 78-95 in Figure 4.7. A similar separation between traveling speed of the BEV efficiency clusters would have been expected to follow suit on class Expressway, especially since this road had a high speed limit, allowing drivers to get a feel for high speed travel in a BEV.
Since all of drivers in the experiment were familiar with the area, due to recruiting through the University, it is also possible that subjects could have been aware of possible penalties for speeding on that stretch of road.
While there is visual evidence from The distributions of the deviation from the speed limit were plotted in Figure 4.8. Both distributions visually had a bell curve structure, which suggested normality.
With both classes of efficient BEV driving representing a bell curve structure, and a number of samples than 40,000 for deviation from the speed limit for each cluster, according to the central limit theorem, the data could be assumed normal. Continuing on the assumption of normality, a t-test was used to determine if the clusters of BEV driving efficiency were statistically different. The test statistic from the t-test resulted in a value of zero, meaning each class of driving efficiency drove the BEV at a different rate of speed when compared to the speed limit. For the sake of comparison to eco-driving in ICE vehi- The second component of eco-driving was defined as aggressive acceleration [9]. One way to determine aggressive driving was to use the safe driving region within a friction circle [84]. The acceleration results from the control runs were plotted in Figure 4.9. The y-axis is the longitudinal acceleration of the vehicle (forward and backward acceleration), while the x-axis is the lateral acceleration (left to right acceleration), measured in meters second 2 . While both control runs had a high degree of acceleration within the safe driving region, the inefficient controls plot from Figure 4.9b had a greater degree of points outside the safe driving region. It was determined that safe, non-aggressive driving behavior results in fewer than 8% of points falling outside the safe driving region, and aggressive driving amounts to more than 10% of points outside of the safe driving region [84]. The efficient control had 4.07% of points outside of the region and the inefficient control had 9.53% of points outside the region. These results pointed to aggressive driving behaviors having an impact BEV on efficiency.  the safe driving region as compared to inefficient BEV drivers. Both clusters' ranges did not expand beyond 10%, which was suggested to be the threshold of aggressive driving [84]. This was most likely due to using the D1 recuperation mode, as this mode applied braking when the driver removed their foot from the gas pedal.
However, inefficient BEV drivers generally had more points outside of the safe driving region, as their median is very close to the third quartile of efficient BEV drivers in Figure   4.10. To test if these acceleration profiles differ, they were first tested for normality.  Thus, acceleration is affecting a BEV driver's efficiency.

Regression of BEV Driving Factors
In an effort to determine the effect that the different factors explored in this research have on BEV efficiency, a hierarchical regression was constructed. To create the regression model, segments of roads were extracted over the course of the experimental route. These segments were randomly generated from the data so that they would only cover uphill, downhill, or flat elevation. The predictor used was the sum of the kWh consumed of the road segment.
Regression 1 is the first model considered in the hierarchal regression and can be seen in the second column of  [12]. The inclusion of the number of points outside the safe driving region of a friction circle was also not significant, as seen in Table   4   The scree plot, depicted in Figure 4.12, shows the expected decrease in eigenvalues as more components are added to the model. Once the scree plot forms an elbow, less variance is explained by adding additional principal components to the model. In Figure 4.12, the elbow formed at the fourth component, which suggested that the data collected could be efficiently explained in a three, four, or five factor model. Table 4.3 displays the numerical eigenvalues from Figure 4.12, while also including the total variance explained by each component added, and the cumulative variance from each component added to the model.
While the scree plot suggested a four factor model, this also was the case for developing the DBS [21], where a three factor model was found to be the most accurate representation of anxiety-based driving. only a three factor model was analyzed. This method of analysis has been seen in other factor analysis experiments using the DBS [87]. Also, a three factor model was able to explain well over half of the variance that exists from the survey answers.
Results from the principal component analysis is outlined in Table 4.4. In the exploratory factor analysis completed to create the DBS, factor loadings greater than 0.3 were considered to be representative of a given factor. Factor loadings, in columns 2-4 in Table   4.4, yielded factor loadings greater than 0.41, in line with the analysis of the DBS [21] and other studies that have done factor analysis on this survey [87]. All the factors are loaded to the same categories as the DBS, with the exception of "I pound on the steering wheel when I'm nervous." This question got loaded with DEF, which measures anxiety related to driving skills. The sample size in this experiment was not as enormous as the DBS, which had over 200 samples [21], or other similar principal component analysis studies which had 147 samples [87]. It is possible that with a larger sample size, all factors would have loaded correctly.

Summary
The results from this analysis show that two clusters of BEV driving efficiency statistically differ. The categories were defined as efficient and inefficient BEV driving. When they were compared to eco-driving in ICE vehicles, both aggressive speed and acceleration were found to affect the rate at which SOC was depleted by the subjects. Anxiety was not found to be a contributing factor to BEV efficiency, as both clusters' answers to the DBS were highly similar.
The final analysis for the effects of efficient and inefficient driving behavior can be found in Table 4.5. The first section of [71]. However, in reality more charging events will happen as these numbers would assume full discharge and full recharge. Three anxiety factors were tested against the efficient and inefficient BEV driving clusters found from hierarchical clustering. Measuring anxiety was an appropriate addition to the study due to range anxiety being key deterrent of market transition to BEVs [62,20].
Multiple forms of anxiety compounded together can be detrimental to an individuals health [22]. This could potentially lead to increased rates of energy consumption for BEV drivers.
Further validation of the results from questionnaire taken during this research was completed using principal component analysis, following similar procedures proposed by the DBS [21] and other research employing the DBS [87]. All questions, except for one, loaded correctly on the three factors of anxiety proposed by the DBS. The three components of anxiety were not found to influence BEV efficiency as the drivers in each of the two clusters answered the questionnaire similarly.
In relation to eco-driving, efficient BEV driving behavior was found to be similar. Since eco-driving was defined as the reduction of aggressive speed and acceleration [9], efficiency clusters were statistically tested against their deviation from the speed limit. Aggressive speed was found to be significantly different between the clusters of efficiency. Speed was expected to have a powerful contribution to efficient BEV driving, as speed is know to exponentially affect the rate that SOC is depleted [12,16,90]. Aggressive use of acceleration was also found to be a factor for BEV efficiency. Using the safe driving region of a friction circle, inefficient BEV drivers statistically had more points outside of the safe driving region as compared to efficient BEV drivers. However, the threshold of 10% for aggressive driving and 8% for non aggressive driving [84] were not determinants in this research, because the BEV was set to a recuperation mode that reduced acceleration. Thus, further analysis of the data included using a t-test to determine if inefficient BEV drivers had more points outside the safe driving region of a friction circle. It was found that the acceleration profiles between efficient and inefficient BEV drivers was significantly different.

Limitations
There are, however, some limitations presented in this research. In terms of driving anxiety, BEV efficiency was studied using the three anxiety factors defined by the DBS. To control range anxiety, drivers were informed that there was more than enoguh energy to complete the experiment. While this allowed for finer analysis of the DBS anxiety factors, this research did not cover the effect that range anxiety could have the clusters of efficiency found. So defining how range anxiety affects the BEV behavioral classes found in this research was not explored.
Lastly, another limitation of this research was that naturalist driving behaviors could not be studied. The vehicle that subjects drove was not their own, the route of travel may not have been one the subjects were accustom too, and a researcher was present during the experimental drive. When studying driving behavior, typically a vehicle is supplied to drivers so that it can be used for daily driving [20]. While participants were given time to adjust the vehicles settings, and were provided a mile travel to become comfortable driving the vehicle, naturalistic driving behavior would be better studied in a vehicle that a subject had for an extended period of time.

Future Work
Results from this research can branch out in various directions, within the scope of BEV energy consumption modeling. To start, the foundation supporting this research can be further expanded in understanding BEV behavioral driving. Those who have studied range anxiety have supplied a BEV to selected drivers for an extended period of time [62].
By supplying a BEV to drivers, behavioral patterns can be studied based on road selection, as well as vehicle comfort settings, to study behavior in uncontrolled environments. With the inclusion of the results from this research, and further research on naturalistic BEV driving behavior, courses could be developed to train drivers to drive BEVs efficiently. In a study tracking one BEV driver, minor interventions helped that driver make better overall route selections [53]. Also, the effect that range anxiety has on BEV efficiency should be explored, as drivers who fit within the classes defined by this research may change when they experience limited range. attempt to better quantify remaining miles given the current charge of the battery and the drivers behavior [92,93]. However, circular range estimation only enables drivers to know how much charge they will use when driving in a straight line, from origin to destination.
Better estimation of energy consumption and the inclusion of BEV behavioral driving can benefit routing problems as well. Most BEV routing problems give a single estimate for the rate of charge used per distance [94,95]. This is also the case for BEV shortest path problems [96,97]. The inclusion of behavioral data will refine these routing techniques, allowing them to model true energy consumption with respect to driving behavior.
BEVs have the potential to develop a more sustainable society. While sales of BEVs continue to grow, these vehicles have not yet captured a prominent share of the automotive market, because potential BEV drivers are concerned with limited range. Manufactures can use the information from this research to develop BEV driving modes that promote efficiency, or learn the current drivers behavior and automatically alter the BEV settings to improve efficiency. Ultimately, this research was able to quantify BEV driving behavior factors that inhibit maximum traveling distance of these vehicles. New BEV drivers can learn to adopt better driving behaviors, and potential BEV drivers can learn from this material, to overcome the hurdle of transitioning from an ICE vehicle to a BEV and help advance the sustainability of society.

ELECTRIC CAR?
Have you always wanted to drive an electric car? Are you curious if it's the same as or different from your car? Do you want to see for yourself how quiet it is?
Well, here is you chance! Come participate in a study that is looking at how we drive EVs. This is a URI research under the supervision of Dr. Gretchen A. Macht which will help understand how people drive electric vehicles, which in turn will help arrange infrastructure better to accommodate the rising use of EVs.
Research volunteers, 18 or older, in possession of a valid American or international driver's license are being sought for an experiment beginning -. This research will be conducted at URI and routes to drive will all be around the Kingston campus.
The experiment will last just over an hour. First, you will take a questionnaire on your perceived driving habits [10 minutes]. Next, you will gain familiarity with the vehicle with a brief training session [10 minutes]. Then you will drive the car on the road on a specified route [50 minutes].
Contact Dan Kowalsky email at dan_kowalsky@my.uri.edu for more information and for recruitment. This research is affiliated with the University of Rhode Island through the MCISE department. This research has been approved by The University of Rhode Island Institutional Review Board.

Want to Test Drive an Electric Car?
Well, here is you chance! Come participate in a study that is looking at how we drive EVs. This is a URI research under the supervision of Dr. Gretchen A. Macht which will help understand how people drive electric vehicles, which in turn will help arrange infrastructure better to accommodate the rising use of EVs.
Research volunteers, 18 or older, in possession of a valid American or international driver's license are being sought for an experiment beginning -. This research will be conducted at URI and routes to drive will all be around the Kingston campus.
The experiment will last just over an hour. First, you will take a questionnaire on your perceived driving habits [10 minutes]. Next, you will gain familiarity with the vehicle with a brief training session [10 minutes]. Then you will drive the car on the road on a specified route [50 minutes].
Contact Dan Kowalsky email at dan_kowalsky@my.uri.edu for more information and for recruitment. This research is affiliated with the University of Rhode Island through the MCISE department. This research has been approved by The University of Rhode Island Institutional Review Board.

Consent Form for Research
The University of Rhode Island

Dear Participant
You have been invited to take part in the research project described below. If you have any questions, please feel free to call Dan Kowalsky or Dr. Macht or Dr. Sodhi, the people mainly responsible for this study.
The purpose of this study is to investigate how driver behaviour changes when combustion engine drivers start using an electric vehicle. All collected information will be stored in Pastore 333 in a locked file cabinet and digital information will be stored in computers in Pastore 254 as locked files. Only the researchers will have access to these files. No audio or video file will be collected. All the information required will be collected directly from the car.

Appendix C -Anxiety Questionnaire
Often times situations occur while people are driving which make them nervous (e.g., weather conditions, heavy traffic, near accidents, etc.). Below is a list of behaviors that may or may not be relevant to you in these situations. Based on your personal experience, please indicate how frequently you perform each of these items when a stressful driving situation occurs which makes