Electric Vehicle Charging Behavior in Existing Infrastructures: A Rhode Island Case Study

The global trend toward a more sustainable future, based on economics and societal behavior, have assisted in making electric vehicles (EV) more attractive to consumers. New technology implemented in EVs continuously improves their range, charging time, and battery capacity. Therefore, the number of EV sales increased significantly within the last few years. In order to handle the demand from the growth in EV sales, the development of a user-orientated distribution of charging stations is needed which requires substantial knowledge about user patterns in charging behavior. Understanding real data from existing charging stations that is analyzed with rigorous statistical methods gives valuable insight for the development of empirical models of charging behavior. In order to initiate this work, a case study approach of public Level-2 charging stations in Rhode Island (RI) were analyzed. Research questions range from how charging stations are being used to which kind of areas influence this behavior and what patterns exist toward calendar dependence. After processing the data, single charging stations were classified into functional areas followed by statistical analysis performed with descriptive statistics, visualizing data, hypothesis testing, clustering, regressive models, and forecasting. Based on the data, there is a strong connection between the total duration of charging events, actual charging time, and the amount of charging events. Not only are chargers utilized differently based on frequency and location, many users use charging stations as parking spots. This pattern exists regardless of charging fees. The charging behavior varies greatly between the different functional areas. Geographical areas seem to have less influence on charging behavior, seemingly more like a mixture of functional areas. Approximately, only about one third of the RI EV drivers are using RI charging stations. There is mainly a decreasing median amount of charges per user, which speaks to either more home charging or larger battery capacity. Areas in which people are working have less charging events on weekends and have a strong peak of charging events in the morning. Areas in which people are spending their free time have the same amount or more charges on weekends and do not have peak times. Timeseries forecasting models found that, both currently and in the near future, there are enough charging stations in RI. However, this does not imply that all the charging stations are in the correct locations, just that the volume of plugs available in RI is sufficient for the current EV charging population. Knowing how people charge their EVs is vital to understanding and implementing a new sustainable transportation infrastructure at a critical time when the monumental paradigm shift has relatively just begun.

the last few years. In order to handle the demand from the growth in EV sales, the development of a user-orientated distribution of charging stations is needed which requires substantial knowledge about user patterns in charging behavior. Understanding real data from existing charging stations that is analyzed with rigorous statistical methods gives valuable insight for the development of empirical models of charging behavior.
In order to initiate this work, a case study approach of public Level-2 charging stations in Rhode Island (RI) were analyzed. Research questions range from how charging stations are being used to which kind of areas influence this behavior and what patterns exist toward calendar dependence. After processing the data, single charging stations were classified into functional areas followed by statistical analysis performed with descriptive statistics, visualizing data, hypothesis testing, clustering, regressive models, and forecasting.
Based on the data, there is a strong connection between the total duration of charging events, actual charging time, and the amount of charging events. Not only are chargers utilized differently based on frequency and location, many users use charging stations as parking spots. This pattern exists regardless of charging fees. The charging behavior varies greatly between the different functional areas. Geographical areas seem to have less influence on charging behavior, seemingly more like a mixture of functional areas. Approximately, only about one third of the RI EV drivers are using RI charging stations. There is mainly a decreasing median amount of charges per user, which speaks to either more home charging or larger battery capacity. Areas in which people are working have less charging events on weekends and have a strong peak of charging events in the morning. Areas in which people are spending their free time have the same amount or more charges on weekends and do not have peak times. Timeseries forecasting models found that, both currently and in the near future, there are enough charging stations in RI. However, this does not imply that all the charging stations are in the correct locations, just that the volume of plugs available in RI is sufficient for the current EV charging population.
Knowing how people charge their EVs is vital to understanding and implementing a new sustainable transportation infrastructure at a critical time when the monumental paradigm shift has relatively just begun. iv ACKNOWLEDGMENTS First, I would like to thank my major professor Dr. Gretchen A. Macht for all the time she put into me and my research. It was always a pleasure working with her even though sometimes it was hard to reach her standards. She made me bring out the best results I could reach. I am thankful that I have been guided by such a strong woman, who had so much influence in my academical development. Even in the most stressful moments she managed to stay positive and bring out the best. I would always take this same path and write my master thesis with Dr. Macht and her amazing SIS Lab.
I also would like to thank Dr. Jyh-Hone Wang. In my undergrad, I never had a statistics class. He was always patient with me when I started learning statistics in an advanced class and taught me so many things I could apply in this thesis.
I would like to thank Dr. Prabhani Kuruppumullage Don for all her time and giving me good advice on how to improve my analysis, and for being on my committee even with non-ideal circumstances.
Furthermore, I would like to thank Dr. Rachel Schwartz who taught me how to handle big data sets and analyze them in an understandable, reproduceable way.
I would like to thank my amazing colleagues from the SIS Lab, it was always fun working with them, sharing good advices and laughing together.
Finally, I would like to thank the Rhode Island Office of Energy Recourses who made this research possible by providing the data.   to nearly 160,000 vehicles, when EVs already hit the 1% market share in the U.S. (Howell, Boyd, Cunningham, Gillard, & Slezak, 2017;Klippenstein, 2017;U.S. Department of Energy, 2017).

LIST OF TABLES
In 2016, the number of registered BEVs was 362 (0.04% market share), and PHEVs 835 (0.09% market share) in Rhode Island (RI) (Auto Alliance, 2016). However, with the initiation of the RI DRIVE rebate program (circa 2016) by the RI Office of Energy Resources (OER), the sales and leases increased by 32 percent. One RI dealership ranked fourth nationally in EV sales behind the first three ranked in California (Faulkner, 2017; Office of Energy Resources, 2018a, 2018b).
New technology implemented in EVs has made transitioning for consumers more feasible by increasing driving range and battery capacity while decreasing charging time (Howell et al., 2017). A prudently designed infrastructure of charging stations is required to effectively serve this growth. EV charging stations are traditionally installed based on property owner interest (ChargePoint, 2018) or algorithms (i.e., simulations, optimizations) with very little to no on user behavior or user expectations, which can lead to increased 2 infrastructure costs further down the road (Chen, Kockelman, & Kahn, 2013;McKerracher, 2016;Wang, Xu, Wen, & Wong, 2013;Wood, Rames, & Muratori, 2018;Xi, Sioshansi, & Marano, 2013 applied to answer specific research. Finally, seasonal forecasting was applied to explore the future evolution of charging events in certain areas. The results and discussion of this analysis can be found in Chapter 4. After that, conclusions can be drawn about the overall usage of the charging stations in Chapter 5. Since there is high potential for additional studies with this data, Chapter 5 also gives a summary about its limitations and future work. The outcome of this research will be published in form of a journal paper(s).

CHAPTER 2 -Literature Review
The popularity of EV's continues to grow and is expected to gain a commanding market share in the near future (Lienert, 2013), as either a bridge to or in lieu of autonomous vehicles (AVs) for private, single users. To make EVs available for the mass market, overall costs have to be reduced (e.g., vehicle and insurance costs, financial incentives), which can happen by improving the transportation system instead of simply investing in larger batteries (Morrow, Karner, & Francfort, 2008). Additionally, lithium-ion battery costs are projected to decrease, reducing production costs thus making EVs even more attractive (Dinger et al., 2010). This can lead to decreasing insurance costs which are high due the possible battery damages.
The capacity, and location of charging stations must be carefully considered in order to effectively support this growth in the EV market, thus the charging station location selection problem becomes an important field of interest (Chen et al., 2013).

EV-MARKET EVOLUTION
Within the last decade the EV-market has grown a lot, but internal combustion engine (ICE) vehicles are still dominating the market with 95% in 2017 (U.S. Department of Energy, 2018). However, there is a significant growth expected within the next years. In 2017 alone, one million EVs were added on the roads globally, making the total global market three million. Europe (yellow), China (blue), and the United States (green) are leading in EV sales right now and are expected to continue growing within the next 20 6 years, see Figure 1 (Bloomberg New Energy Finance, 2017). Based on projections, this implies that in the U.S. alone there will potentially be one million EVs on the road by 2020. Additionally, the number of EV models is growing based on manufacturers' sustainability goals. There were over 10 new models available in 2017, 41 in total and the number is growing based on support from Volvo, BMW, Volkswagen, Telsa, and more(ChargePoint, 2017); within a matter of a few years, there should be model for every use and personal preference (ChargePoint, 2017). Unfortunately, some states have special restrictions for EVs; Tesla for example is not available in every state, due to dealer franchise laws (Gatti, 2017).
The causes of this fast-paced growth is, but not limited to, the new technology implemented in EVs and the fast falling battery costs, which will make EVs become price competitive. Trends like car sharing, and vehicle-as-a-service (e.g., leases with a monthly mile limit) models will influence the market as well, much earlier in adoption and market penetration timing than autonomous vehicles (Bloomberg New Energy Finance, 2017 Investing in the charging infrastructure also count for investing in the EV market in general, see Figure 2. EVs are dependent on charging, and charging stations only justify themselves when there are enough EVs. Investments in charging infrastructure must be made before they are needed to initiate this loop, what causes a utilization gap at the beginning of the implementation, see Figure 3. That implies that there have to be stations, 9 even if they are not used yet; charging infrastructure needs to be available, for people to even consider buying an EV. However, there will be a market pull after EVs reach a certain percentage of market share, and prior investments will no longer be needed. After enough investments in the EV-market has been done it will continue to grow by its own, within a reinforcing loop, see Figure 2 (Meadows, 2009;Wood, Rames, & Muratori, 2018).
Figure 3: Correlation of charging infrastructure requirements (Wood et al., 2018) Charging stations are currently available at three different types: (1) Level-1 takes up to 22 hours charging time and is mostly used residentially (home charging), through a convenient household outlet; (2) Level-2, with approximately four to eight hours charging time, is used primarily for commercial charging and is the most common in RI, but they

EVS IN RHODE ISLAND
Rhode Island has visions about the future development and the City of Providence, as its capital, has set a statement in writing about self-improvements. There are three attributes they want to identify with, mentioned in the Comprehensive Plan of Providence are, they want to be more "green", with a healthy natural environment and a sustainable city design.
Also, Providence wants to be more "efficient", a fiscally sound city, providing high-quality and cost-effective services. In their guiding principles, the City of Providence also mentions "sustainability", regarding climate change and the uncertainties within the oil market. Therefore, Providence wants to promote alternatives to the conventional traveling patterns (Taveras, 2014) Regarding the chosen attributes, a growing trend in transportation with EVs has to be considered. Transportation is the costliest energy sector in Rhode Island (Office of Energy  (Elsworth, 2016).

CHARGING STATION LOCATION MODELS
There are different approaches to examining the problem of charging station location selection. Existing studies use mathematical models, studies which use linear programming, analyzing driving and charging behavior to work on this problem. Primarily, researchers create mathematical models to simulate the problem in order to solve for potential solutions. There are multi-objective planning models (e.g., gas-station demands, power grid infrastructure) to layout charging station distributions, in some studies they also determine scheduling of charge and operation costs (Chen, Kockelman, & Kahn, 2013;McKerracher, 2016;G. Wang, Xu, Wen, & Wong, 2013;Wood, Rames, & Muratori, 2018;Xi, Sioshansi, & Marano, 2013). Simulation-optimization models determine where to locate EV charging stations in order to maximize the use for privately owned EVs (Xi et al., 2013). There are also studies which utilize multi-objective planning models to improve the transport system efficiency, as well as improve the grid system operations (Luo, Zhu, Wan, Zhang, & Li, 2016).
Other studies use linear programming to include electricity price variation, the capability of EVs to charge and discharge when desired, called Vehicle-to-Grid technology (Kristoffersen, Capion, & Meibom, 2011;Srivastava, Annabathina, & Kamalasadan, 2010). Further research on vehicle-to-grid technology provided evidence that this approach was not viable due to high infrastructural costs and significantly shortened lifespan of EV batteries by the higher number of charge/discharging cycles (Göthel & Bräul, 2012;Mullan, 2012), but there are still stakeholders (ChargePoint, 2017). Certain station location selection models are based on existing optimization routines or heuristics that can find charging locations based on reducing queuing times via prediction of existing data from non-EV vehicles (Chen et al., 2013;De Weerdt, Gerding, Stein, Robu, & Jennings, 2012;Worley, Klabjan, & Sweda, 2012). Yet, these algorithms or models still barley consider EV users preferences, behavioral patterns, and functional areas analysis. Additionally, EV charging is different than refueling an ICE vehicle, therefore studies should focus on EV driving patterns.
Another approach to the problem is by analyzing driving behavior from the EV driver: analyzing their driving, parking, and charging patterns. This level of research has been attempted by making test drives or tracking fleet vehicles (Smart & Schey, 2012;Speidel & Bräunl, 2014). In Australia, they observed EV driving and charging behavior of a fleet of eleven EVs at 23 Level-2 Charging stations. They assessed the state of charge (SOC) before and after the charging events, charging time, time the vehicles are plugged in at the station and energy consumption. No categorizations of public level-2 changing stations exist based on location (Speidel & Bräunl, 2014). Additionally, fleet vehicles are operated differently than consumer-owned EVs, so their behavior may not transition to predictions of charging station locations. Another study, performed in the U.S., with 2903 privately owned Nissan LEAFs looked at SOC and the number of charging events; they only differ 15 between home charging and "away-from-home location" (Smart & Schey, 2012 Plugs. The input variables a user of the projection tool can insert are: "Number of vehicles to support", "Vehicle Mix" (Percentage BEV and PHEV), "Support of PHEVs", and "percentage of driver with access to home charging" (U.S. Department of Energy, 2018).
The EVI-Pro tool forecasts quantity and type of charging plugs, there are no behavioral recommendations and no functional area considerations for charging station locations.
Therefore, personal use of EVs are considered limited with respect to charging behavioral patterns even with current growth, use, and installations throughout the literature.

CHAPTER 3 -Methodology
This chapter describes the methods that have been used to analyze the data and discusses the motivating factors for this specific approach. The flow of this chapter starts with describing the data and its processing, followed by the classification of functional areas, and ended up with the specific methods of statistical analysis which have been applied. At the beginning it appeared that there was data of 57 charging stations but while verifying the locations one was identified with only two test charges; this station will not be part of the analysis. Another station appeared twice in the data set, because the owner and the name changed, the station data was classified as one.

CLASSIFICATION OF FUNCTIONAL AREAS
Every city and town in RI classifies their areas into different functional districts by their zoning ordinance to control the land use. The following items are covered in a RI Zoning Ordinance: site layout requirements, requirements for structure characteristics, permitted use, and procedural matters (Atash, 2017). Every RI city and town publishes their zoning map and explanations online for anyone to find a location's zone classification and its regulations. The location for every charging station was established with respect to their functional zone and added to the dataset. For example, the City of Providence has 20 charging stations available, with 40 charging ports in total, located and distributed in different functional districts classified by the City of Providence Zoning Ordinance committee. This zoning ordinance information will be examined to see if charging stations in different functional areas are utilized differently. A detailed analysis of these types of zones/functional areas for charging stations occurred for the City of Providence.
The City of Providence is regulated by their zoning ordinance and represented in the zoning maps as eight districts which individually have more subdivisions (City of Providence Zoning Ordinance, 2017). There are "Residential Districts" (yellow to brown, Figure 6) with different dimensional standards and housing types. "Commercial Districts" (pink to red, Figure 6) with medium-scale to intense commercial use and design standards. The "Downtown District" (gray, Figure 6) has a special focus in the Comprehensive Plan of Providence as it is a mixed-use district with special regulations. physical improvements has to be regulated and approved through development plan review in accordance with the provisions of this area. Furthermore, there are two "Institutional Districts" (blue, Figure 6) in Providence: one with a special focus on healthcare and one with a special focus on education. The "Industrial Districts" (violet, Figure 6) incorporate light to heavy intensity industrial uses, some of which are mixedused and includes also residential or commercial use. The "Waterfront District" (turquoise blue, Figure 6) incorporates residential, commercial and industrial uses with special restrictions regarding the waterfront. "Open Space and Public Space" (green, Figure 6) are summarized as one zone. Open spaces include parks, wetlands, floodplains, cemeteries, and conservation areas. Public Spaces are areas for public buildings and facilities such as parks and recreation areas or schools. The last zone incorporates the "Special Purpose Districts" which have intense focus on certain areas of interest to the City of Providence (City of Providence Zoning Ordinance, 2017a). Districts and four are located in Industrial Districts (see Figure 6).
This functional classification has been assessed for every city and town with charging stations in the dataset. However, there are slight differences in the classification of the districts, because every city and town can name their own districts.
This required reconciling these differences between word choices and zoning types for a more unified approach. Examples of these differences and how they were reconciled are articulated as the following: (1) Barrington has one charging station in a "Business Districts", for commercial and retail activities (Town of Barrington Zoning Ordinance, 2003). This is assumed to be an equivalent to a "Commercial District", since Barrington has not classified any "Commercial Districts"; this appears to be a different designation for the same kind of district. The same classification adjustment was made for Narragansett, Warwick and West Greenwich (City of Warwick Zoning Ordinance, 2018; Town of Narragansett, 2012; Town of West Greenwich Zoning Ordinance, 2017).
(2) In Lincoln there is one charging station in a "Manufacturing District". Since there is no district classified as an "Industrial District", it assumed that manufacturing is an industrial use and therefore now classified as an industrial area (Town of Lincoln Zoning Ordinance, 2015).
(3) In the Town of Smithfield, home to Bryant University with two charging stations, the town has not classified any districts as "Institutional Districts".
Yet, a University is definitely institutional, that is why the charging stations at the Bryant University are clustered as Institutional (Smithfield Zoning Ordinance, 2018).
(4) The T.F. Green Airport in Warwick has two charging stations, it is classified as a "Intermodal District". Since the use of an airport is other to any of the other mentioned districts, the T.F. Green Airport has keeps this classification (City of Warwick Zoning Ordinance, 2018).
(5) All "Open Space" and "Public Space" are as in the plan of Providence grouped together.
The locations of the charging stations in all other cities and towns could be classified into the previously used areas: "Commercial District", "Downtown District" (only applies to Providence), "Industrial District", "Institutional District", "Intermodal District" (only applies to Warwick), "Open Space and Public Space", "Residential  (2) there is a monotonic relationship between the two variables (Groß, 2010). Spearman's Correlation was applied in Chapter 4.1 since all three variables meet both assumptions.
In Figure 7, the monotonic relationships between each of the variables are illustrated.
For the comparison of proportions, a two-sample proportion test was ran using Minitab.

Figure 7: Relationship between Total Duration (TD), Charging Events (CE) and
Charging Time (CT) With hierarchal clustering, findings can be grouped within a tree-structured cluster dendrogram generated with R or Minitab. There are different options that can be chosen: complete linkage (i.e., similarity of the furthest pair); single-linkage (i.e., similarity of the closest pair); group average (i.e., similarity between groups); centroid similarity (i.e., iteration merges the most similar central point independent to each other, which requires special methods that explicitly consider stochastic dependence (Groß, 2010). Since, the data was collected at any time a charging event happens and summarized into various time groups (e.g., years, month, days…), they can be treated via timeseries.
If the data is summarized in fixed time intervals (in this case years or month) there can be a linear dependence, called autocorrelations with lag , which can be tested with an autocorrelation function. The data summarized as days has significant noise, which makes timeseries analysis with this data inaccurate; thus, it will not be assessed at the daily level. With the Box-Pierce and Ljung-Box test, based on an autocorrelation function (ACF), the data can be tested as to whether the observations can be treated as independent variables or not. If they are independent, special time series analysis is not necessary. Box-Pierce and Ljung-Box tests were performed with a 95% confidence level throughout this thesis (Groß, 2010 (Brockwell & Davis, 2002;Groß, 2010).
To find values ( , ) which could give a good fit the ACF and PACF of the given time series can be plotted; this study performed it in R. The procedure performed for ARMA ( , ) models which is stationary data, with differencing non-stationary data can become stationary which is done prior plotting ACF and PACF. An alternative to this procedure is the R function " . ()" which shows the probably most suitable model (Brockwell & Davis, 2002;Groß, 2010).
After that the models with good fit can be tested with the R-function " ()".
The model with the smallest AIC (Akaike Information Criterion) is the preferred model, the AIC is a measure of quality for the fit of a model in the maximum likelihood principle. An ARIMA ( , 1, ) model works with differenced data, to reveal the trend again, the drift can be included, this can also improve the AIC value. After that the fit with the chosen model can be proofed visually with the R function " ()". The standardized residuals should look kind of like white noise; in the ACF of residuals there should be no value be over the blue dotted line after lag 0; the p-values for Ljung-Box statistics should be over the blue line. If all these conditions are fulfilled the model has a good fit (Brockwell & Davis, 2002;Groß, 2010).
If a good fitting model is found the values can be predicted with the R function " ()", the periods (in this case number of month) which should be predicted ahead can be chosen. Also, the forecasting can be visualized with the function " ( ())", the number of periods ahead can be chosen and if all data should be included.
For seasonal forecasting a SARIMA (seasonal ARIMA) can be applied. A periodic seasonal pattern has to be noticeable which is in correlation with a constant time period , this is not the case in the given dataset. Otherwise a SARIMA ( , , ) ( , , ) model could have been applied (Groß, 2010).

CHAPTER 4 -Results & Discussion
This chapter seeks to answer the research questions from Chapter 1. All the findings are out of the provided dataset of 50 public Level-2 stations and analyzed with the mentioned tools in Chapter 3. The fee operated charging stations are within the data but will be part of some additional analysis. This section provides an overview about the RI data in general and compares it to nationwide trends. Firstly, all the energy savings (i.e., energy used, greenhouse gas

GENERAL USAGE OF CHARGING STATIONS
[GHG] savings, and gasoline savings) and time factors (i.e., duration plugged in, charging time, and parked after fully charged) are examined with descriptive statistics for all the charging stations. Table 1 shows the mean value, the standard deviation (SD), the minimal value (Min), the maximal value (Max) and the total of all charging events.
The measured factors are the energy consumption of the charging events in kilowatthours (kWh), GHG savings in kilograms (kg) due to the gasoline savings in gallons, and the total duration that the vehicles have been plugged in at the station. In addition, the actual charging time and the time an EV driver parked at the charging spot after they were fully charged is also measured and documented in common time format. The RI charging stations total use has saved the emissions from around 64 ICE passenger 29 vehicles driven for 1-year since their installation (US Environmental Protection Agency,

2017).
A notable finding is that the mean charging time almost equals the mean parking time. In earlier studies on charging behavior, users were spending only about 10% of the time for charging out of the total time plugged into a charging station, while in our dataset the mean charging time was very similar (1:55:52) (Speidel & Bräunl, 2014).
This could show a shift of usage trends; this will be explored later.  The total duration is classified at the time the vehicle is plugged in during a charging event. Figure 9 shows the total duration divided by the total time installed the year 2017, except for stations 24, 37, 49 and 50 which were newly implemented in 2017.
A charging station could be occupied 100% of the time but based on the lack of charging traditionally from 11pm to 5am a realistic utilization per charging station is 75% ( Figure   9 station 37). Only three stations have a utilization over 50% (Figure 9 gray line)., with the maximum being stations 37 at 78.32% of the time. The median utilization is 6.40%     Null hypothesis (H0) is that the station is not significantly different. One station in an industrial area charges $0.1/kWh, it is expected the station is not significantly different to the others since they do not pay anymore after they are fully charged since 35 58% of the users at this station leave within 30 min and 42% stay longer. This station is not significantly different to the general non-fee use; with a p-value of 0.5, it fails to reject the H0 hypothesis.
The other two stations are in a commercial area located next to each other, the first 4-hours are free and after that the charge is $1/hr. As expected, this cost model appears to prevent users from occupying the station for too long. At these stations 70% and 71% use the charging stations just to charge and 30% and 29% stay longer than 30-minutes after they are fully charged. This is better than the other stations, but it could also be due to the area in which the stations are located. With an p-value under 0.001 for both tested, this rejects the H0 hypothesis for the stations with a $1/hr fee after the fourth hour, they are significantly different from the general use non-fee chargers. These two stations are not significantly different to each other (p = 0.55), but the they are significantly different to the other payed station with $0.1/kWh ( p < 0.001 for both).
It was found that one charging station with fee is used like they are generally used in RI. The other two charging stations with a different fee model (fee after 4-hours) behave differently. But that is not necessarily only due to the fee, it can also be due to the functional area in which the charging stations are located. To test if this is the reason for the different behavior, further analysis will be made in Chapter 4.2.1.

4.2 USAGE WITH LOCATIONS AS A FACTOR
The location seems to have strong impact on the user behavior at charging stations.
In this section discusses Research Question 2: Does the type of areas in which the charging stations are located influence the patterns of charging behavior?

FUNCTIONAL AREAS
A map was created to get an idea where the charging stations are located and their functional areas. The RI map in Figure 12 shows each of the 50 charging stations as dots, colored according to their area type and their size representing how many charging events occurred in 2017.    Table 2 are descriptive statistics about the total duration of charging events, charging times, and amount of charges per station in functional areas. The mean value, the standard deviation, and the maximum value are given; no minimum value is shared since it is always zero. It is quite notable that these values vary drastically between functional areas. In the cluster dendrogram in Figure 17 paired with  Comparing the amount of charging events by county from Figure 18, Providence has the most charging events, percentage wise even more than charging stations. All other counties have percentage wise less charging stations as charging events.     can they predict future usage trends?
Working with time series requires caution, because they usually are dependent on time. Looking at the number of charging events (n) per day ( Figure 24) there is already a trend noticeable, but for working with time series this form is not ideal because there is a lot of noise due to daily variation. Summarizing the data further can be helpful to be less affected by outliers.  To analyze the similarities and differences in functional areas, they have been summarized in weekdays. In Figure 26 you can see how different it looks in the certain areas per weekday. It is noticeable that downtown, industrial, and institutional areas have more charging events from Monday to Friday and less at the weekends, this could be attributed to people working in these districts. Commercial and intermodal areas do not show strong differences between weekends and the rest of the week, people go shopping every day, and also using the airport at any day. Residential areas are more inconsistent, because this is usually the origin of drives. Open space areas show a different pattern, there are more charging events during the weekends compared to other days, possibly because people visiting these areas more often at weekends.   Charging station users are primarily using the RI charging stations mostly between 7am and 7pm with a peak at 8am (Figure 29). In comparison with the nationwide trends, published from ChargePoint 2016, the usage in RI shows slight differences. The top 8 charging times from 7am to 2pm are very similar, however in RI there is another uptick peak at 5pm where in the rest of the U.S. the amount of charging events is constantly falling from 2pm (McKerracher, 2016). This pattern is similar to that in a study from Australia but 1 hour delayed, peak time at 9am comparing to 8am in RI, after 2pm constantly falling (Speidel & Bräunl, 2014).  The clustering of the functional areas by daytime to validate observations it can be seen in Figure 31's dendrogram. Figure 30 illustrates a strong relationship with 94% between commercial and open space areas, which link with the intermodal area at 76% similarity. Downtown and institutional areas have 97% similarity, industrial and residential areas 94%, this both groups link with 80% similarity. All areas have 61% similarity to each other. These groupings are slightly different than previous clustering groups with intermodal left the larger group to joining commercial and open spaces.

57
For seasonal forecasting a periodic seasonal pattern has to be noticeable which is in correlation with a constant time period (s). Figure 32 shows the charging events per month, it is noticeable that there is always a dip in February, but this is not enough for seasonal forecasting. When divided in functional areas ( Figure 33) there is also in no recognizable seasonal trend. Additionally, the RI counties are not showing any seasonal trends either as seen in Figure 34.    The null hypothesis for Box-Pierce and Ljung-Box is that observations can be treated as independent. When the amount of charging events is summarized by month, the p-values for Box-Pierce and Ljung-Box are both p < 0.001 which rejects the null hypothesis; thus, the observations cannot be treated as independent. Therefore, an ARIMA model will be fitted, the analyzed data summarized as month is non-stationary, the mean value from the given data is changing over time based on the Augmented Dickey-Fuller test " . ()" in R (p = 0.08977; fail to rejectH0; data is nonstationary). In the data there is a noticeable upwards trend, an ARIMA ( , 1, ) model will be used. The maximum likelihood estimators AR-coefficient ( ) (order of autoregressive terms) and the MA-coefficient ( ) (number of lags on the MA component) are found with looking at the ACF and PACF. Therefore, first the data has been differenced and tested again with the Augmented Dickey-Fuller test (p = 0.01; rejectH0; data is stationary). In Figure 38, it is noticeable that there are no significant values after lag 0 in both the ACF and the PACF, which implies that an ARIMA (0,1,0) could be a good fit. The same model has been advocated by the R function " . ()".   Improving the forecasting by putting in more data is tried, there for the charging events are now summarized as weeks. The null hypothesis for Box-Pierce and Ljung-Box is that observations can be treated as independent, both (p < 0.001) rejected the null hypothesis, the observations cannot be treated as independent. Therefore, also as sums of weeks an ARIMA model can be fitted. The analyzed data seemed to be stationary and confirmed via the Augmented Dickey-Fuller test (p = 0.03678; reject 0 , data is stationary). The mean value from the given data is changing over time and visually there is a perceivable trend. Looking at the ACF in Figure 41, the trend is also noticeable with many significant values, the PACF has significant values at lag one and two. The ARIMA (1,1,1) model has an AIC value of 2265.56 which was the lowest for this model between all the different models tested. The assumption that charging behavior is strongly dependent on the functional area has been further affirmed, regarding the timeseries data for charging behavior per weekday and daytime. The geographical influence on the charging behavior seems to be less, seemingly more like a mixture of functional areas. Predictions and forecasting of further demand on charging stations could be made. Still, future trends on charging behavior are strongly influenced by several factors. This is the reason why the predictions can only be seen as approximate trends.

CHAPTER 5 -Conclusion
The purpose of this research was to gain valuable insight into the usage of charging stations using Rhode Island as a case study. Each research question was intended to give special insight into different aspects of how EV drivers use public charging stations.
The first research question explored the overall use associated with these RI public be drawn with linear regression models. It has been found that only about one third of the RI EV drivers are using RI charging stations. Additionally, timeseries forecasting models were performed with the data and found that currently there are enough charging stations in RI. Currently, RI is in the 'utilization gap' with respect to number of charging stations and usage, however the data was limited in terms of volume. As years pass and EVs become more ubiquitous, more data is required for a more accurate prediction.
This research has given an overview about the charging behavior in Rhode Island.
The outcomes of this work can help the RI Office of Energy Resources and the RI 70 Department of Transportation to plan further action on EV infrastructure. This research could also be the corner stone of many following research projects in this field as it is the first of its kind. The research outcome can be compared to other states in the U.S., where nationwide trends could be analyzed and maybe even predicted. Knowing how people charge their EVs is vital to understanding and implementing a new sustainable transportation infrastructure at a critical time when the monumental paradigm shift has relatively, just begun.

LIMITATIONS
The given data set has a high potential for being utilized in future research. The scope of this particular research was to give an overview about the data and initial trends found within the dataset.
The first limitation to this work is that it is highly location based; implying that all charging stations are within the state of Rhode Island. It is unclear at this time if how people utilized charging stations in Rhode Island is similar or different to Florida, Missouri, California, or even Washington. Additional studies should expand the scope of this work toward understanding both nationally and internationally the differences in charging behavior.
The second limitation is the dataset itself. The data is public, government sponsored charging stations. How this data and conclusions interact with private charging stations and residential charging stations is still an area for opportunity in researching for human variability in transportation.

71
Another limitation is that this work is single-user based. There were assumptions that the person charging at the station was the same person and that the vehicles charged at that station were the same throughout time. Further expansions of this research can exist where clarifications of this can be used to explore in depth the variability in charging behavior based on vehicle type, person driving, and even links to driving behavior. Along with being a single-user based focused study, the limitations in this area is that it is not a closed-system. The fleet EV research was very specific to medium duty vehicles, but the system remained the same with just the drivers being the variable.
In this study, all these facets in the system were not controlled and explored assuming all with equal value and weight. Future work can surely expand upon this work with additional data for a more comprehensive understanding at the individual user-level.

FUTURE WORK
This analysis is the cornerstone for the following, planned research projects.
Seasonality could be investigated more in depth. Potential questions are, but not BEV, 50% PHEV, so it is assumed that RI will get to this number by 2020 (ChargePoint, 2017). Another study says that PHEVs will only play a role in EV adoption until 2025 and thus, after 2030 they will almost be gone. As the battery capacity grows, so does the range, resulting in a higher amount of higher range vehicles estimate over the next few years (Howell et al., 2017). With this information the EVI-Pro tool calculates that RI provides a lot more charging plugs than needed currently and in the near future as RI is still in the utilization gap. Also, the electric driving range influences the results significantly; out of the EVI-Pro tool the number of needed charging plugs does not 74 increase a lot despite a growing number of users due to the higher electric drive range (U.S. Department of Energy, 2018).
More research on potential influencing factors needs to be explored (e.g., battery evolution and electric driving range, user expectations, linkage between charging time and hour of the day, vehicle type). Especially the linkage between home charging and battery capacity needs to be invested; a vital question is, "Do EV drivers charge more at home if the driving range is larger?". Typically, Level-1 charging is used at home, do people decide to switch to Level-2 charging at home instead if the batteries have more capacity and take longer to charge? How do patterns of human behavior change these factors? These factors could be further investigated and input into a model in order to predict charging behavior. The outcomes of these studies could all be implemented into projection tools like EVI-Pro.
These are just a few ideas of projects that could be done with this data; embrace the possibilities.

Data Analysis Master Thesis in R Roxana Voss
June 2018

Electric Vehicle (EV) Charging Behavior in existing Infrastructures
This is an R Markdown document to understand the processed statistics of the data. It is a trial version, which keeps record of possibly usable statistics for the research. Updated Data can easily be read in and the same analysis can be performed automatically.