1. Trang chủ
  2. » Tất cả

Effect of time of day and day of the week on congestion duration and breakdown: a case study at a bottleneck in salem, NH

10 3 0
Tài liệu đã được kiểm tra trùng lặp

Đang tải... (xem toàn văn)

THÔNG TIN TÀI LIỆU

Thông tin cơ bản

Tiêu đề Effect of Time of Day and Day of the Week on Congestion Duration and Breakdown: A Case Study at a Bottleneck in Salem, NH
Tác giả Eric M. Laflamme, Paul J. Ossenbruggen
Trường học Plymouth State University
Chuyên ngành Traffic and Transportation Engineering
Thể loại research paper
Năm xuất bản 2017
Thành phố Salem
Định dạng
Số trang 10
Dung lượng 0,93 MB

Các công cụ chuyển đổi và chỉnh sửa cho tài liệu này

Nội dung

Effect of time of day and day of the week on congestion duration and breakdown A case study at a bottleneck in Salem, NH Q3 ww sciencedirect com 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 2[.]

Trang 1

Original research paper

Effect of time-of-day and day-of-the-week on

congestion duration and breakdown: A case study

at a bottleneck in Salem, NH

Q3 Eric M Laflammea,*, Paul J Ossenbruggenb

aDepartment of Mathematics, Plymouth State University, Plymouth, NH 03264, USA

b

Department of Mathematics and Statistics, University of New Hampshire, Durham, NH 03824, USA

h i g h l i g h t s

 Features of recurrent congestion and recovery are analyzed via regression models

 Probability of recurrent congestion increases between AM- and PM-rush periods

 Effect of time-of-day on congestion duration depends on the day-of-the-week

a r t i c l e i n f o

Article history:

Available online xxx

Keywords:

Stochastic models

Ordinary least squares regression

Binary logistic regression

Congestion duration

Breakdown

a b s t r a c t

This work uses regression models to analyze two characteristics of recurrent congestion:

breakdown, the transition from freely flowing conditions to a congested state, and duration, the time between the onset and clearance of recurrent congestion First, we apply a binary logistic regression model where a continuous measurement for traffic flow and a dichoto-mous categorical variable for time-of-day (AM- or PM-rush hours) is used to predict the probability of breakdown Second, we apply an ordinary least squares regression model where categorical variables for time-of-day (AM- or PM-rush hours) and day-of-the-week (MondayeThursday or Friday) are used to predict recurrent congestion duration Models are fitted to data collected from a bottleneck on I-93 in Salem, NH, over a period of 9 months

Results from the breakdown model, predict probabilities of recurrent congestion, are consistent with observed traffic and illustrate an upshift in breakdown probabilities between the AM- and PM-rush periods Results from the regression model for congestion duration reveal the presences of significant interaction between time-of-day and day-of-the-week

Thus, the effect of time-of-day on congestion duration depends on the day-of-the-week This work provides a simplification of recurrent congestion and recovery, very noisy processes

Simplification, conveying complex relationships with simple statistical summaries-facts, is

a practical and powerful tool for traffic administrators to use in the decision-making process

© 2017 Periodical Offices of Chang'an University Publishing services by Elsevier B.V on behalf of Owner This is an open access article under the CC BY-NC-ND license (http://

creativecommons.org/licenses/by-nc-nd/4.0/)

* Corresponding author Tel.: þ1 603 724 5336

E-mail addresses:emlaflamme@plymouth.edu(E.M Laflamme),pjo@unh.edu(P.J Ossenbruggen)

Peer review under responsibility of Periodical Offices of Chang'an University

Available online at www.sciencedirect.com

ScienceDirect

journal homepage:w ww.elsevier.com/locat e/jtte

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129

http://dx.doi.org/10.1016/j.jtte.2016.08.004

2095-7564/© 2017 Periodical Offices of Chang'an University Publishing services by Elsevier B.V on behalf of Owner This is an open

access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Trang 2

1 Introduction

Traffic congestion costs American drivers billions of dollars in

wasted fuel and loss of productivity Among all types of

congestion, recurrent congestion, congestion occurring every

day, accounts for roughly half of the congestion experienced

by Americans (U.S Department of Transportation, 2016)

Furthermore, these recurring congestion events on highways

account for 40% of the total delay, more than the delay from

construction and traffic incidents (non-recurring events)

combined (Paniati, 2003)

Clearly, recurrent congestion is a problem and methods

must be developed to alleviate the situation But, before

expensive construction or remedial measures are employed, it

is critical that administrators identify trends/characteristics

of recurring congestion on their roadways As stated byRao

and Rao (2012), identifying characteristics of congestion

events will serve as a guide to administrators to help them

choose appropriate measures to mitigate such congestion

Along these same lines,Hao et al (2007)stated that in order

to control or alleviate congestion effectively, researchers

must investigate its key features

So, what are these characteristics of recurrent congestion,

and the key features of these events pursued by researchers?

Key features may refer to the underlying causes of congestion

For example, both high flow during peak-hours (Downs, 2004)

and physical structures such as bottlenecks (Ban et al., 2007)

have been shown to be responsible for recurrent congestion

Other features include congestion dynamics, the evolvement

of the congestion event such as queue propagation (Newell,

1993) and shock-wave analysis (Bertini and Cassidy, 2002;

Kerner, 2004) Other common features of congestion are the

stochastic nature of breakdown, the transition from freely

flowing traffic to a congested state, and the factors that

trigger these events (Elefteriadou et al., 1995) Other

researches are devoted to measures/metrics that quantify

their extent, duration, and intensity of congestion (Shaw,

2003) Such measures include level-of-service, congestion

duration (the time between onset of breakdown and

clearance of congestion event, or, alternatively, the time that

the travel rate indicates congested travel on a segment),

travel time index, etc In all cases, no matter the

characteristic, researchers aim to better understand

recurrent congestion and identify the underlying dynamics

of the process

With these works in mind, authors focus on two particular

characteristics of recurrent congestion: breakdown and

duration Authors will only focus on recurrent events, and,

going forward, “congestion” will always means “recurrent

congestion” As mentioned above, breakdown refers to the

transition from a freely flowing traffic state to a congested

state This is the standard definition and commonly used

phrase of breakdown given by Kerner (2009) Of course,

identifying breakdown relies on how we define “freely

flowing” and “congested” The precise criteria for identifying

freeflow and congestion, and thus how we define

breakdown, will be discussed later in Section2.4 Congestion

duration refers to the time between the onset of breakdown

and clearance of a congested event This too relies on how

we define a congested traffic state This definition and the precise criteria used to identify congestion duration will be discussed later in Section2.5

Why did authors choose these particular congestion characteristics? First, congestion duration is chosen because,

as a“time-based” measure of congestion, it is in keeping with the common perception of the congestion problem (Rao and Rao, 2012), yet the explicit investigation of congestion duration's statistical characteristics has attracted only limited attention in the literature Vlahogianni et al (2011)

state that the dynamics of congestion duration may contain useful information about intraday traffic operations and should be further explored Second, breakdown was chosen because, despite receiving ample treatment in the literature (Elefteriadou et al., 1995; Persaud et al., 1998), it remains a controversial topic That is, while the stochastic nature of traffic breakdown has been verified (Elefteriadou et al., 1995), the mechanism or trigger of breakdown is still mysterious It

is our goal to gain some insight into these two aspects of recurrent congestion

In this work, authors will use two separate regression an-alyses where breakdown and congestion duration are used as respective response variables We will then introduce explanatory variables derived from traffic stream data to identify underlying factors associated with both breakdown and duration Ideally, from our models, identifying significant predictors of breakdown and congestion duration will lead to

a better understanding of recurrent congestion

The remainder of this paper is organized as follows Sec-tion 2 presents authors' materials and methods including statistical (regression) model forms, the raw traffic stream data from authors' collection site, preprocessing procedures, data aggregation, and extraction of the requisite variables for model fitting Section 3 presents authors' results from model-fitting and interpretation of these results Lastly, section4is the conclusion of authors' study

2 Materials and methods

2.1 Statistical models: binary logistic regression model for breakdown probability

A traditional approach to identifying probability of breakdown

is the use of a generalized linear model (GLM) GLMs have the basic form for random response variable Y

gðmiÞ ¼ Xib (1) where the mean value mi is given by E(Yi), g($) is a smooth monotonic“link” function, Xiis the ith row of model matrix X, andb is a vector of unknown model parameters GLMs assume that the Yiare independent and Yi~ some exponential family distribution To model breakdown probability, the authors let

Y represent the traffic state where Y¼ 1 denotes congested traffic and Y ¼ 0 denotes freely flowing traffic Then, mi is interpreted as the probability,p, of Yitaking on the value of one, Pr(Y¼ 1), and the authors use the canonical link function with the form of Eq.(1)

gðpÞ ¼ ln p

1 p



(2)

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 3

Such models are known as binary logistic regression

models For further details regarding logistic models and

GLMs, the reader is directed toHosmer et al (1989)

Several analyses have been performed in which probability

of breakdown is identified as a function of flow

measure-ments.Athol and Bullen (1973)suggested that the expected

time to breakdown is a declining function of flow Lorenz

and Elefteriadou (2001)illustrated the probabilistic nature of

breakdown and showed empirical evidence that probability

of congestion increases with increasing flow rate Other

studies have similarly shown that congestion breakdown

occurs with some probability at various flow rates, that

breakdown is, in fact, a stochastic process (Dong and

Mahmassani, 2009a,b; Elefteriadou et al., 1995; Evans et al.,

2001).Persaud et al (1998)explored the relationship between

flow and the probability of breakdown empirically, by visual

assessment Ossenbruggen (2016), used stochastic

differential equations to predict breakdown probability

based on average flows for discretized time-of-day periods

Following these analyses, traffic flow (volume) will be used

as the primary predictor of breakdown

In addition to flow, and because breakdown probability is

likely influenced by time-of-day, a time variable will be

introduced into authors' regression model In the spirit of

simplification, our time-of-day measure will be a categorical

variable distinguishing between two time sectors: AM-rush

(7 a.m.e2 p.m.) and PM-rush (2 p.m.e7 p.m.) These periods

were not chosen arbitrarily, but based on several factors First,

and most simply, the vast majority, more than 90%, of

congestion events occur between 7 a.m and 7 p.m Second,

within these 12 h, the AM-rush and PM-rush periods were

distinguished/identified based on a visual comparison of

average flows throughout the day (Fig 1) This figure

illustrates how average flows within each period are similar

in terms of magnitude and variability, yet dissimilar to one

another That is, average flows from 2 p.m to 7 p.m

(PM-rush) are higher and have more variability as compared to

flows from 7 a.m to 2 p.m (AM-rush) Third, identification

of these periods was further based on a previous analysis

where a piecewise linear function was fitted to daily average

flows In this investigation, there was a distinct and sharp

increase in average flow at 2 p.m., lending credence to the choice of 2 p.m as a break point between AM- and PM-rush periods Furthermore, this analysis revealed a nearly flat trend in average flow from around 7 a.m up to 2 p.m., while flows after 2 p.m were more volatile and higher until about

7 p.m These estimated trends reinforced the AM- and PM-rush periods identified visually InOssenbruggen (2016), this piecewise linear modeling approach and these exact time periods were used to calibrate a stochastic differential equation model used to predict traffic breakdown In fact, these same AM- and PM-rush time periods were identified

as significant predictors of breakdown Fourth, in a study of both recurrent and non-recurrent congestion events,

Hallenbeck et al (2003) identified 3 p.m as a natural break between midday and afternoon peak periods While this work does not perfectly agree with our choice of 2 p.m as break point, it is very similar and we feel our choice is justified based on trends in average flows Also, this analysis supports our decision to simplify time-of-day in terms of a categorical variable with just a few variables Fifth, in an analysis of recurring congestion, Mazzenga and Demetsky (2009) identified AM- and PM-peak periods between 5 a.m.Q1

and 8 a.m and between 3 p.m and 7 p.m., respectively

Again, this does not agree with our periods exactly, but supports the use of simplified time-of-day categories The authors are confident that the time periods capture the recurring events and distinguish the morning and afternoon breakdowns As a final note, the authors acknowledge that time-of-day likely serves as a proxy for more complicated,

“lurking,” secondary traffic characteristics (aggressive driving, weaving, etc.) that cannot be extracted from the traffic stream data

InFig 1, vertical lines represent breaks between AM-rush and PM-rush periods Fig 1illustrates the homogeneity of traffic flows within each time period

In addition to a flow main effect and a time sector dichot-omous variable, the interaction between the time and the continuous flow variable will be included in our model spec-ification This will allow for a possible shift (change in inter-cept) and a change in slope (actually, a change in the“log odds”) between the two time sectors

Fig 1e Average flow and variability by time-of-day

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 4

So, for our breakdown model, the probability of breakdown

Pr(Y¼ 1), denoted p, may be expressed as follow

p ¼ expðb0þ b1qþ b2dþ b3qdÞ

1þ expðb0þ b1qþ b2dþ b3qdÞ (3)

where q is the flow, d is the dichotomous variable

dis-tinguishing AM- or PM-rush hours, qd is the interaction

vari-able The procedure for identifying our binary response

variable, freeflow or congested state, and the associated

pre-dictor variables will be discussed later in Section2.4

In terms of statistical methodology,Persaud et al (2001)is

especially relevant as the authors use logistic regression

models based on 3-lane, 1-min flow averages to predict

congestion events Also,Ossenbruggen (2016)uses a logistic

form for breakdown probability as a component of a

stochastic differential equation to explain the congestion

process To our knowledge, no statistical analyses have been

performed using our methodology and exact combination of

predictors

2.2 Statistical models: ordinary least squares regression

for congestion duration

To analyze congestion duration, we use an ordinary least

squares (OLS) regression platform Specifically, our response,

W is congestion duration in minutes, a continuous variable

measured from the onset of breakdown to the clearance of the

congestion event The authors chose two explanatory

vari-ables in model: time-of-day and day-of-the-week Since most

drivers have seen some association between time-of-day and

the presence/severity of the congestion, considering

time-of-day as a potential predictor of congestion duration makes

intuitive sense Furthermore, since flow follows a distinct

pattern daily, time-of-day may be considered a simple proxy

for traffic flow InTebaldi et al (2002), day-of-the-week was

used as a primary predictor of traffic flow in their

hierarchical regression models In the literature, however,

day-of-the-week was not commonly used to distinguish

traffic stream characteristics In fact, most analyses used

data collected from weekday traffic interchangeably

Germane to this analysis, Falcocchio and Levinson (2015)

summarized trends in recurrent congestion duration by both

time-of-day and day-of-the-week The authors found

end-of-week traffic more severe, a perspective supported by the

own driving experience at the collection site and from an

initial data investigation

Under the OLS regression format, a congestion duration

length W then has the following form

W¼ b0þ b1xþ b2dþ b3xdþ e (4)

where x is the categorical variable distinguishing

day-of-the-week, xd is the interaction variable, ande  Nð0; seÞ and

in-dependent The procedure for identifying congestion duration

and the associated predictor variables will be discussed later

in Section2.5 Such a model (continuous response with two

categorical predictors) could also be analyzed using an

analysis of variance (ANOVA) format In fact, with such

dichotomous predictors, the two model forms and their

corresponding hypothesis tests produce equivalent results

The study of non-recurrent congestion duration (incident duration) has received ample treatment in the literature An-alyses byGarib et al (1997), Giuliano (1989), andSullivan (1997)

used lognormal distributions to describe freeway incident duration Conditional models for incident duration have been pursued by Jones et al (1991), for example.Nam and Mannering (2000) used hazard-based models to find the likelihood that an incident will end in the next short time period given its continuing duration Similarly,Stathopoulos and Karlaftis (2002) used a probabilistic log-logistic functional form to describe incident durations A number of works have used OLS regression models to investigate the association between incident duration and certain traffic stream variables (Garib et al., 1997; Gomez, 2005) To our knowledge, however, no statistical analyses have been performed using our methodology to analyze recurrent congestion duration

To fit the regression models, the authors use real-world traffic stream data collected by the New Hampshire (NH) Depart-ment of Transportation at a collection site in Salem, NH, along the northbound lane of I-93 just north of an off-ramp, exit 1, and just south of an on-ramp A bottleneck occurs here as, immediately north of this location (downstream), I-93 is physically constricted from three to two lanes (Fig 2) In addition to this physical bottleneck, traffic volumes at this site exceed 100,000 vehicles per day (VPD) which far surpass the 60,000e70,000 VPD that the roadway was designed to accommodate (U.S Department of Transportation, 2016)

Because of this heavy, daily flow at the bottleneck, recurrent congestion occurs here As stated byBrilon et al (2005), such sites are ideal for the collection of congestion data

Data collection occurred between April 1 and November 30,

2010 During this time, side-fire radar devices intermittently measured traffic at irregular but frequent time periods about

1 min apart Data observations (raw data) consist of the following measurements: vehicle counts, average speed, oc-cupancy, and speed (spot speed) of individual vehicles observed over the interval Since incidents of recurrent congestion are the sole focus, data obtained during weekends and holidays are omitted from the analysis Also, because of scheduled maintenance as well as unscheduled“gaps” where the radar devices stopped collecting data (some lasted for several days), many other days were omitted In the end, 128 complete days of quality data were retained Without any

Fig 2e Illustration of collection site structure along northbound lanes of I-93

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 5

records of non-recurrent events during the collection period,

the authors assume that all congestion events observed on

these days are recurrent

Next, because radar data were collected over very short,

irregular time intervals, these measurements were aggregated

into uniform intervals of 15 min 15-min intervals are

rec-ommended to ensure“stable” flow rates that are especially

suitable for macroscopic/speed-flow analyses (Smith and

Ulmer, 2003; TRB, 2010) Specifically, harmonic averages were

calculated from flow and spot speed observations within each

15-min interval to produce an aggregated flow rate (q) and

aggregated speed (u) in units of vehicles per minute (vpm) and

miles per hour (mph), respectively (Daganzo, 1997) Thus, for

each day, aggregation yields ut and qt (speed and flow,

respectively) where t represents time-of-day with t¼ 1, 2, …,

96.Fig 3illustrates the speed and flow aggregates produced

for one week in April Also, Fig 3illustrates the stochastic

nature of breakdown, how high flows typically, but do not

necessarily, result in breakdown (and high, sustained flows

are more likely to result in breakdown) Lastly, the authors

note that the figure includes the flow profiles during a

weekend (the third and fourth mounds) where, as is

typically the case, no breakdowns occur This supports the

decision to remove weekend days from the analysis of

recurrent congestion

InFig 3, black dots indicate speeds less than 50 mph, the

critical threshold speed Notice that breakdown corresponds

to sustained, high traffic flows on the first, fifth, and seventh

days Also, notice that no breakdown occurs on the third

and fourth days, which is a weekend

2.4 Extraction of variables for binary logistic regression

model

From the speed and flow aggregates, the variables required by

the regression models can be easily extracted

First, to identify a response variable Y to distinguish

be-tween congested and freely flowing traffic (to identify a traffic

breakdown), a fixed speed threshold u*is typically chosen to

identify the transition between them (Banks, 2006; Brilon

et al., 2005; Geistefeldt and Brilon, 2009; Habbib-Mattar et al.,

2009; Lorenz and Elefteriadou, 2001; Yeon et al., 2009) No

standard approach exists for identifying u*, but based on

bimodal speed aggregates, u*¼ 50 mph was chose This

tran-sitional threshold is similar to those of Brilon et al (2005),

Geistefeldt and Brilon (2009), and Lorenz and Elefteriadou

(2001), Yeon et al (2009), who used fixed values of 47 mph (75 kph), 50 mph, 43 mph (70 kph), and 56 mph, respectively So, for some time t, ut> u*and utþ1> u*indicates breakdown at time t and thus Yt¼ 1 Among others,Lorenz and Elefteriadou (2001) used a similar approach to identify breakdown from traffic stream data If, on the other hand, ut > u* and

utþ1> u*

, no breakdown occurs at time t and thus Yt ¼ 0

This simple rule was applied to create a response Ytfor each time t in the data

Because the GLM form assumes the independent response values, we must select an independent set of Y values for model fitting First, since congested observations are typically separated by long time intervals and periods of freeflow, the authors may assume that the congested responses are, by their rare nature, independent of one another Next, the au-thors take a random sample of uncongested responses This random sampling“removes” the serial correlation from these data and destroys any time structure that exists in the aggregated values The authors then combine the two sets of responses, the congested and sampled uncongested values, to form an independent set of Y values suitable for our model form

Next, for each response Ytconsidered, the corresponding flow aggregate is observed at time t So, when breakdown occurs at time interval t, qt represents the flow observed immediately prior to breakdown (commonly called a “break-down flow”).Kondyli et al (2013), who pursued a variety of breakdown identification methods (speed-based, occupancy-based, and volume-occupancy/correlation-based), identified these “breakdown flows” in a similar situation When breakdown does not occur at time t, qt represents a traffic flow under freely flowing conditions

Lastly, for each response Yt, a corresponding categorical variable for time-of-day is simply identified based on time markers retained from the aggregation process

Despite starting with a time-ordered, correlated set of aggregated values, we have extracted a set of data that con-tains no temporal structure By considering an independent set of response values (see above) and their corresponding flow and time-of-day values, we have extracted an indepen-dent set of data that is suitable for a regression analysis

2.5 Extraction of variables for duration model

For our duration modeling, the authors must identify a response variable W that represents the time between onset of

Fig 3e Speed and flow aggregates for one week (April 1, 2010eApril 7, 2010)

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 6

breakdown and clearance of breakdown, the time required by

the system to clear a congestion event As previously

dis-cussed, time of breakdown is first identified via transition

from sustained speeds above to below the threshold u* Then,

clearance is similarly identified via transition from sustained

speeds below to above the threshold ut< u* and utþ1 > u*

indicate a clearance at time t Then, for each congestion event,

a duration W is simply calculated by the time between onset

and clearance This definition agrees withRao and Rao (2012)

who stated that the duration of congestion can be determined

by measuring reduced travel speeds over a period of time

Other more complicated techniques exist for duration

estimation by Elefteriadou et al (2011a,b) who used a

wavelet transform method to identify the start and end time

of congestion event occurrence, but this simpler method

was applied

For all congestion events, a histogram of congestion

du-rations is given below (Fig 4) This figure reveals a very

right-skewed distribution with some very long congestion events

The median congestion duration is 120 min

Lastly, based on time and date markers retained from the

raw data, time-of-day and day-of-the-week categorical

vari-ables were identified for each congestion event These

cate-gorical variables were matched with each duration amount to

create a dataset to be used for our duration regression model

Because there were one or two congestion events per day,

which were separated by a period of freeflow, each duration

measurement was assumed to be independent In an initial

data investigation, no temporal dependence between

dura-tions was observed Thus, this data is appropriate for our

regression model

3 Results and discussion

Based on the variables extracted from the I-93 data, the

lo-gistic and OLS regression models for breakdown and duration,

respectively, model fitting via maximum likelihood

estima-tion was performed using R statistical software

3.1 Binary logistic regression model for probability of breakdown

The binary logistic regression model was fit to the following data: flow, the continuous explanatory variable; time sector designation, the dichotomous categorical explanatory vari-able; and congestion state, the binary response Based on in-dividual p-value analyses (using a 5% significance level) and Chi-squared tests for change in deviance, both the flow and the dichotomous time variables were found to be highly sig-nificant explanatory variables in the prediction of breakdown

The interaction term, however, was not found to be significant and thus omitted from the model Thus, in terms of the probability curve for the two time sectors (AM- and PM-rush), (1) there is a significant shift, and (2) there is no significant change in trend Residual deviances indicate no significant evidence of lack-of-fit of the model

Based on our fitted models, an “S” shaped curve repre-senting probability of breakdown for observed flows bp was produced for both time periods (Fig 5) 95% confidence bands were included to illustrate uncertainty associated with probabilities at certain flow values

95% confidence bands (light and dark gray bands) are included to illustrate variability associated with logistic curves

First, the authors analyze the time-of-day variable The fitted coefficient associated with this dichotomous variable was found to be highly significant (z¼ 6.750, p z 0) The odds ratio, given by dOR¼ expðbb2Þ, is the relative increase in odds of congestion when going from AM-rush (d ¼ 0) to PM-rush (d¼ 1) The fitted parameter estimate of bb2¼ 2.17 corresponds

to an odds ratio of 8.76, or about an 8-fold increase in the odds

of breakdown when going from AM-to PM-rush When considering the 95% confidence interval for bb2, (1.54, 2.80), this odds ratio is between 4.66 and 16.44 in 95% confident, clearly a significant (non-zero) increase in odds between time periods

From the plot of the logistic curves for AM- and PM-rush pe-riods (Fig 5), this change in odds is observed in the clear shift

Fig 4e Frequency distribution (histogram) for congestion

durations

Fig 5e Fitted logistic curves for AM- and PM-rush time periods

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 7

between the two curves While some overlap between

confidence bands occurs at high flows, this is simply

because the uncertainty associated with the AM-rush period

(dark bands indicate 95% confidence for AM-rush period)

increases at high flows where few observations exist

It can be concluded that, for any flow, the probability of

breakdown during the PM-rush is higher than that during the

AM-rush (that PM-rush traffic behavior is generally more

likely to be congested) Observed traffic behavior supports this

finding as congestion is much more common during the

PM-rush period than during the AM-PM-rush In fact, of the

conges-tion events considered, about 80% are observed during the

PM-rush period Certainly, flow is the most important driver of

congestion, but our result suggests that time-of-day may play

a role The authors suspect that driver behavior during these

periods may contribute to this discrepancy, that PM-rush

commuters likely drive more aggressively In any event, the

results warrant further investigation and analysis Lastly,

steps to minimize congestion duration for the AM-rush, say,

may not be effective for the PM-rush

This result illustrates the limitation of a single, fixed

ca-pacity value as prescribed by the Highway Caca-pacity Manual

(HCM) According to the HCM guidelines, the freeway capacity

for a single lane of traffic is 2000 vehicles per hour (vph) Since

our traffic observations were taken across three traffic lanes,

the HCM capacity for these three lanes is 6000 vph or 100

ve-hicles per minute (vpm) At that flow, for flows of 100 vpm, the

probability of breakdown is somewhat low (around 20%)

during the AM-rush period, while probability of breakdown

during the PM-rush period for the same flow is about 70%

Since the interaction term was not significant and omitted

from our model (b3was not found to be significantly different

from zero), it is assumed that both the AM- and PM-rush

pe-riods share a common trend That is, the change in odds of

breakdown for a one unit increase in flow is the same for both

AM- and PM-rush periods This shared change in odds is the

odds ratio associated withb1, given by dOR¼ expðbb1Þ From our

fitted model results, the coefficient estimate of bb1¼ 0.03

cor-responds to an odds ratio of 1.031 Thus, for both periods, the

probability of breakdown increases 3.1% for every one

addi-tional vehicle per minute From the plot above, both periods

have probabilities that progress at the same rate relative to

flow (Fig 5)

From an administrator's point of view, there may be a

practical application of these results Most simply, results

from the breakdown model suggest that measures to remedy

recurrent congestion be focused on PM-rush periods, when

congestion is more likely For example, hard shoulder running

may be implemented during these congestion-prone hours to

mitigate expected congestion This practice is a proven

tech-nique for congestion mitigation in Northern Virginia

(Mazzenga and Demetsky, 2009), and the extension of its use is

recommended (Bauer et al., 2004) Furthermore, this practice

has been shown to effectively ease congestion without

increasing accident rates (ITS International, 2013) Or,

administrators may opt to introduce variable speed limits

(VSLs) during these congestion-prone hours Such VSLs are

used extensively in Europe with great success (Mazzenga

and Demetsky, 2009) As a final example, administrators

may implement ramp metering during the PM-rush It has

been shown that ramp metering is effective at limiting the number of vehicles entering a freeway and can help to prevent recurrent congestion (Texas A&M Transportation Institute, 2016)

3.2 Ordinary least squares regression model for congestion duration

The least squares regression model was fit to the following data: time-of-day, a dichotomous explanatory variable dis-tinguishing AM- and PM-rush time sectors; day-of-the-week, a categorical explanatory variable; and congestion duration, the continuous response Initially, the categorical variable repre-senting day-of-the-week contained a level (category) for each weekday, Monday through Friday However, after several model-fitting exercises, it was determined that most days were not significantly different from one another In fact, based on an analysis of variance, data collected from Monday, Tuesday, Wednesday, and Thursday are nearly identical

However, it was found that MondayeThursday data was significantly different from Friday data, so a dichotomous variable was used in lieu of a categorical variable for day-of-week Also, initial models suggested a lack of normality among the residuals To remedy this, a transformation of the response was pursued Because duration times are highly right-skewed, and because variance seems to increase with duration, a log-transformation was applied (Fig 4) Therefore, the slightly modified model is as follow

logðWÞ ¼ b0þ b1d1þ b2d2þ b3d1d2þ e (5) where d1is a dichotomous variable distinguishing AM- and PM-rush periods, d2is a dichotomous variable distinguishing Fridays from all other days (d2¼ 1 if Friday, d2¼ 0 if Monday, Tuesday, Wednesday, or Thursday), d1d2 is the interaction term Similarly, Garib et al (1997) used linear regression models to predict the log of duration of traffic incidents observed from California freeways (non-recurrent events)

While the model fits the data surprisingly well (R2suggests the model form explains about 67% of the total variation observed in log(Duration) measurements), a “better” model was not pursued While attempting to find a model that maximizes the explained variation is a useful exercise, our aim is to identify significant explanatory variables and their implications on congestion duration

Of primary importance from our fitted result is the statis-tical significance of the interaction term (t ¼ 3.277, p-value¼ 0.00189) Both main effects, time-of-day and day-of-the-week, are also found to be statistically significant (t¼ 4.323, p-value ¼ 7.15e-05; t ¼ 4.455, p-value ¼ 4.61e-05, respectively), although main effects are typically ignored in the presence of interaction That is, if two terms interact, changes in both explanatory variables will have an effect on the response outcome In our case, the effect on the mean outcome, log(Duration), from a change in time-of-day de-pends on the level of day-of-the-week This effect is most easily seen in the plots below (Fig 6) Interaction lines calculated from the fitted regression model are included

The two plots above represent the same phenomenon from two different perspectives: grouping the data by day-of-the-week (Fig 6(a)) and time-of-day (Fig 6(b)) In either case, a

dis-1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 8

ordinal interaction is marked by the clearly non-parallel

interaction lines.Fig 6(b) is probably the most interpretable

as the moving along the x-axis represents transition from

AM-to rush periods Here, when going from AM-to

PM-rush periods (x-axis grouping), the change in log(Duration)

depends on day-of-the-week For MondayeThursday, there

is an increase in log(Duration), while there is a decrease in

log(Duration) for Fridays Analysis of the fitted parameters

allows us to precisely quantify this effect After converting

back to the original scale of measurement (duration in

minutes instead of log(Duration)), calculations reveal that

for MondayeThursday traffic, transitioning from AM-rush

period to PM-rush period results in more than a 2-fold

increase (273%) in duration On the other hand, for Friday

traffic, transitioning from AM-to PM-rush periods results in a

59% decrease in duration Similarly, fromFig 6(a), there is a

distinct change in log(Duration) when going from Monday to

Thursday to Fri categories, and the magnitude of the change

depends on the time-of-day For the AM-rush, this change

(from Monday to Thursday to Fri.) is a steep increase in

log(Duration) For the PM-rush period, this change is

increasing as well, but more modestly

Results from our duration modeling are a bit surprising

The authors can speculate that Friday“evening” commutes

may be occurring earlier in the day Since our analysis did not

identify a significant number of congestion events in the early

afternoon, though, this idea was abandoned Perhaps the

evening commute is spread across a wide time span and

re-sults in fewer long congestion events This conjecture would

agree with the results presented in this analysis In any event,

this result may warrant a more day-specific analysis of

congestion events If our results are confirmed, and

Mon-dayeThursday traffic experiences longer congestion events,

perhaps administrators should implement congestion

miti-gation techniques during these days or possibly extend these

mitigation techniques later into the evening, beyond

tradi-tional“rush hours” As mentioned above, these methods to

remedy such congestion may be hard shoulder running, VSLs,

or ramp metering

Lastly, the regression/model fitting diagnostics suggest

that the regression assumptions are all met (that our model

specification is appropriate for this data) Plots of residuals versus fitted values show constant variance and normality plots show the residuals to be normally distributed Also, there was no indication of correlated errors, further evidence that the duration response values are completely uncorrelated and suitable for the OLS format

4 Conclusions

Recurrent congestion is a complicated process that is likely triggered by a variety of interconnected conditions That said, understanding the frequency and duration of these events is critical to highway administrators and decision-makers The goal of this study is not to create a real-time forecasting tool, but rather to identify model structures that reveal the importance, driving mechanisms of the congestion process

Such structures will contain information that is easily inter-preted and understood

In our binary logistic regression model to predict proba-bility of breakdown by flow, results show a distinct shift by time-of-day For any flow, the PM-rush period has higher probability of breakdown than the AM-rush period Overall, based on the parameter estimate associated with the time-of-day, there is an estimated 8-fold increase in the odds of breakdown when going from AM-to PM-rush So, while flow is the primary predictor of congestion, our result indicates that time-of-day is likely a factor as well From an administrator's point of view, this may suggest that measures to remedy recurrent congestion should be focused on PM-rush periods

As previously mentioned, these remedial measures may include hard shoulder running, variable speed limits (VSLs), or ramp metering

By considering time-of-day as a simple proxy for secondary traffic characteristics, the result from our logistic regression analysis may suggest that driver behavior affects congestion and roadway capacity That is, one could speculate that there are more aggressive drivers during the PM-rush hour, and these aggressive drivers are prone to excessive weaving, etc

With more detailed data, future work could investigate these

Fig 6e Boxplot for congestion duration (a) MondayeThursday and Friday traffic by time-of-day (b) AM- and PM-rush

periods for day-of-the-week

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 9

specific, underlying factors rather than time-of-day

categories

In our duration model, an ordinary least squares regression

model is used to predict congestion duration using

time-of-day and time-of-day-of-the-week categorical variables After an initial

investigation, it suffices to use dichotomous variables for

time-of-day, AM- or PM-rush, and day-of-the-week,

Mon-dayeThursday or Friday The primary result from this model

is the statistically significant interaction term This suggests

that when going from AM-to PM-rush periods, the change in

congestion duration depends on day-of-the-week In fact, this

change is increasing for MondayeThursday and decreasing

for Fridays This surprising result may suggest the need for

more day-specific analysis of congestion events Ultimately,

administrators may choose to focus congestion mitigation

techniques (hard shoulder running, VSLs, ramp metering, etc.)

during MondayeThursday, evening commute hours

This work represents an empirical study on breakdown

and delay at a known bottleneck in Salem, NH Our respective

model specifications successfully identify significant

pre-dictors that shed light upon the congestion process A

limi-tation of this analysis, however, is that our conclusions are

restricted to the one location from which data was collected

That is, our results are not transferable to other locations,

even those very near our location or in other directions Data

from additional bottlenecks will allow for transferable results,

a general model that could explain the congestion process

across diverse locations That says, the results here are

justi-fied and illuminate some congestion phenomena that are

likely not site-specific

Uncited reference

Texas A and M Transportation Institute, 2015

Q2

r e f e r e n c e s

Athol, P.J., Bullen, A., 1973 Multiple ramp control for a freeway

bottleneck Highway Research Record 456, 50e54

Ban, X., Chu, L., Benouar, H., 2007 Bottleneck identification and

calibration for corridor management planning Transportation

Research Record 1999, 40e53

Banks, J., 2006 New Approach to Bottleneck Capacity Analysis:

Final Report UCB-ITS-PRR-2006-13 California PATH, Institute

of Transportation Studies, University of California, Berkeley

Bauer, K., Harwood, D., Hughes, W., et al., 2004 Safety effects of

narrow lanes and shoulder-use lanes to increase capacity of

urban freeways Transportation Research Record 1897, 71e80

Bertini, R.L., Cassidy, M.J., 2002 Some observed queue discharge

features at a freeway bottleneck downstream of a merge

Transportation Research Part A: Policy and Practice 36 (8),

683e697

Brilon, W., Geistefeld, J., Regler, M., 2005 Reliability of freeway

traffic flow: a stochastic concept of capacity In: 16th

International Symposium of Transportation and Traffic

Theory, College Park, 2005

Daganzo, C., 1997 Fundamentals of Transportation and Traffic

Operations Pergamon, Oxford

Dong, J., Mahmassani, H.S., 2009a Flow breakdown, travel

reliability and real-time information in route choice behavior

In: 18th International Symposium on Transportation and Traffic Theory, Hong Kong, 2009

Dong, J., Mahmassani, H.S., 2009b Flow breakdown and travel time reliability Transportation Research Record 2124, 203e212

Downs, A., 2004 Still Stuck in Traffic: Coping with Peak-Hour Traffic Congestion The Brookings Institution Press, Washington DC

Elefteriadou, L., Kondyli, A., Washbum, S., et al., 2011a

Proactive ramp management under the threat of freeway flow breakdown Procedia-Social and Behavioral Sciences 16, 4e14

Elefteriadou, L., Martin, B., Simmerman, T., et al., 2011b Using Micro-Simulation to Evaluate the Effects of Advanced Vehicle Technologies on Congestion 2009-06 University of Florida Transportation Research Center, Gainesville

Elefteriadou, L., Roess, R.P., McShane, W.R., 1995 Probabilistic nature

of breakdown at freeway merge junctions Transportation Research Record 1484, 80e89

Evans, J., Elefteriadou, L., Gautam, N., 2001 Determination of the probability of breakdown on a freeway based on zonal merging probabilities Transportation Research Part B:

Methodological 35, 237e254

Falcocchio, J.S., Levinson, H.S., 2015 Road Traffic Congestion: a Concise Guide Springer International, Switzerland

Garib, A., Radwan, A.E., Al-Deek, H., 1997 Estimating magnitude and duration of incident delays ASCE Journal of Transportation Engineering 123 (6), 459e466

Geistefeldt, J., Brilon, W., 2009 A comparative assessment of stochastic capacity estimation methods In: 18th International Symposium on Transportation and Traffic Theory, Hong Kong, 2009

Giuliano, G., 1989 Incident characteristics, frequency, and duration

on a high volume urban freeway Transportation Research Part A: General 23 (5), 387e396

Gomez, N.M., 2005 Incident Duration Model and Secondary Incident Causation Model Based on Archived Traffic Management Center Data (Master thesis) Vanderbilt University, Nashville

Habbib-Mattar, C., Abishai, P., Cohen, M.A., 2009 Analysis of the breakdown process on congested freeways Transportation Research Record 2124, 58e66

Hao, Y., Xu, T., Sun, L., 2007 Analysis and control of recurrent traffic congestion on urban expressway In: International Conference on Transportation Engineering, Chengdu, 2007

Hallenbeck, M.F., Ishimaru, J.M., Nee, J., 2003 Measurement of Recurring Versus Non-Recurring Congestion: Technical Report

WA-RD 568.1 Washington State Department of Transportation, Seattle

Hosmer, D.W., Lemeshow, S., Sturdivant, R.X., 1989 Applied Logistic Regression John Wiley, New York

ITS International, 2013 Traffic Monitoring and Hard Shoulder Running Available at: http://www.itsinternational.com/

categories/detection-monitoring-machine-vision/features/

traffic-monitoring-and-hard-shoulder-running/ (Accessed 1 May 2016)

Jones, B., Janssen, L., Mannering, F., 1991 Analysis of the frequency and duration of freeway accidents in seattle

Accident Analysis and Prevention 23 (4), 239e255

Kerner, B.S., 2004 The Physics of Traffic Springer, New York

Kerner, B.S., 2009 Introduction to Modern Traffic Flow Theory and Control Springer, New York

Kondyli, A., Elefteriadou, L., Brilon, W., et al., 2013 Development and evaluation of methods for constructing breakdown probability models ASCE Journal of Transportation Engineering 139 (9), 931e940

Lorenz, M., Elefteriadou, L., 2001 Defining freeway capacity as function of breakdown probability Transportation Research Record 1776, 43e51

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

Trang 10

Mazzenga, N.J., Demetsky, M.J., 2009 Investigation of Solutions to

Recurring Congestion on Freeways Virginia Transportation

Research Council, Charlottesville, 2009

Nam, D., Mannering, F., 2000 An exploratory hazard-based

analysis of highway incident duration Transportation

Research Part A: Policy and Practice 34 (2), 85e102

Newell, G.F., 1993 A simplified theory of kinematic waves in

highway traffic I: general theory II: queuing at freeway

bottlenecks III: multi-destination flows Transportation

Research Part B: Methodological 27 (4), 289e303

Ossenbruggen, P., 2016 Assessing freeway breakdown and

recovery: a stochastic model ASCE Journal of Transportation

Engineering 142 (7), 04016025

Paniati, J.F., 2003 Using Intelligent Transportation Systems (ITS)

Technologies and Strategies to Better Manage Congestion

ITS Joint Program Office, Washington DC

Persaud, B., Yagar, S., Brownlee, R., 1998 Exploration of the

breakdown phenomenon in freeway traffic Transportation

Research Record 1634, 64e69

Persaud, B., Yagar, S., Tsui, D., et al., 2001 Study of

breakdown-related capacity for a freeway with ramp metering

Transportation Research Record 1748, 110e115

Rao, A.M., Rao, K.R., 2012 Measuring urban traffic congestion: a

review International Journal for Traffic and Transport

Engineering 2 (4), 286e305

Shaw, T., 2003 NCHRP Synthesis 311: Performance Measures of

Operational Effectiveness for Highway Segments and

Systems Transportation Research Board, Washington DC

Smith, B.L., Ulmer, J.M., 2003 Freeway traffic flow rate

measurement: an investigation into the impact of the

measurement time interval Journal of Transportation

Engineering 129 (3), 223e229

Stathopoulos, A., Karlaftis, M.G., 2002 Modeling duration of

urban traffic congestion ASCE Journal of Transportation

Engineering 128 (6), 587e590

Sullivan, E.C., 1997 New model for predicting freeway incidents

and incident delays ASCE Journal of Transportation

Engineering 123 (4), 267e275

Tebaldi, C., West, M., Karr, A.F., 2002 Statistical analyses of

freeway traffic flows Journal of Forecasting 21 (1), 39e68

Texas A&M Transportation Institute, 2015 Traffic Gridlock Sets New Records for Traveller Misery Available at: http://

mobility.tamu.edu/ums/media-information/press-release/

(Accessed 1 July 2015)

Texas A&M Transportation Institute, 2016 Ramp flow control e Urban Mobility Information Available at: http://mobility

tamu.edu/mip/strategies-pdfs/active-traffic/technical-summary/Ramp-Flow-Control-4-Pg.pdf (Accessed 1 June 2016)

TRB, 2010 Highway Capacity Manual Transportation Research Board, Washington DC

U.S Department of Transportation, 2016a Bridging Multiple Communities: New Hampshire's I-93 Improvement Project

Available at:http://www.fhwa.dot.gov/construction/accelerated/

wsnh06.pdf(Accessed 1 May 2016)

U.S Department of Transportation, Federal Highway Administration, 2016b Reducing Recurring Congestion

Available at: http://ops.fhwa.dot.gov/program_areas/reduce-recur-cong.htm(Accessed 1 May 2016)

Vlahogianni, E., Karlaftis, M.G., Kepaptsoglou, K., 2011 Nonlinear autoregressive conditional duration models for traffic congestion estimation Journal of Probability and Statistics

2011, 798953

Yeon, J., Hernandez, S., Elefteriadou, L., 2009 Differences in freeway capacity by day-of-the-week, time-of-day, and segment type ASCE Journal of Transportation Engineering

135 (7), 416e426

Eric M Laflamme is a professor of Mathe-matics at Plymouth State University in Ply-mouth, NH He received a Master's degree in Applied Statistics from Cornell University and a PhD in Mathematics from the Uni-versity of New Hampshire His areas of research are transportation and extreme value theory related to climate change projections

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80

Ngày đăng: 24/11/2022, 17:43

TỪ KHÓA LIÊN QUAN

TÀI LIỆU CÙNG NGƯỜI DÙNG

TÀI LIỆU LIÊN QUAN

🧩 Sản phẩm bạn có thể quan tâm

w