미리보기
서지주요정보
Uncovering and modeling inter-city human mobility: Focusing on highway and railway travels = 도시간 인간 이동에 대한 이해와 모델링: 고속도로와 철도 통행을 중심으로
서명 / 저자 Uncovering and modeling inter-city human mobility: Focusing on highway and railway travels = 도시간 인간 이동에 대한 이해와 모델링: 고속도로와 철도 통행을 중심으로 / Jiho Yeo.
발행사항 [대전 : 한국과학기술원, 2020].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8036329

소장위치/청구기호

학술문화관(문화관)B1층 보존서고

DGT 20005

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

The dissertation discusses the long-term evolution patterns of inter-city travel networks and the demand prediction problem in the inter-city level. By using railway and highway travel data from 1977 to 2016, the dissertation verifies that the changes in inter-city travel networks have close relationships with changes in population. As people have concentrated on large cities over the years, the residents of large cities are interacting with other cities more frequently than in the past. This strengthened inter-city connectivity is the evidence that the life space of people is gradually growing, centered on hub cities. The results also show that the direction of evolution varies by transportation modes. The role of the highway and railway became divergent as two travel networks have evolved over the decades. This result implies that highways evolved to serve short-distant travels while railways developed to handle long-distant trips. On the other hand, the dissertation also provides the traffic demand prediction model in the inter-city level. The model predicts daily Origin-Destination traffic demands between cities. The novel Graph Convolutional Network is proposed to consider both spatial and temporal dependencies. The stratified framework, which divides the heterogeneous O-D graph into multiple subgraphs and trains them separately, is utilized to capture the large heterogeneity of traffic demand. The performance of the model is tested by using the inter-city travel data of Korean highways from 2015 to 2019, showing the best performance compared to the state-of-the-art model. The dissertation shows that convergent and interdisciplinary research presents a variety of perspectives on investigating human mobility. Analysis of the network through graph theory provided various perspectives for observing the characteristics of human mobility, and the attributes of these graphs contributed to the development of high-precision prediction models along with the deep learning. Grafting of various academic fields and the use of domain knowledge in transportation are significantly helpful in interpreting, understanding, and modeling human mobility.

본 학위논문은 지난 40년 동안의 철도와 고속도로을 통한 이동 정보를 수집하여 도시 간 통행 네트워크 장기적인 변화를 관찰하고, 도시간 통행 수요를 예측하는 모델을 다룬다. 철도와 고속도로 모두에서 분석 기간 동안 도시 간 통행 불균형이 점차 심화되고 있으며, 이는 인구분포의 도시 집중화로 인한 인구의 양극화 현상과 밀접한 연관이 있었다. 도시간 통행 네트워크의 장기적인 변화는 교통 인프라의 중심인 대도시의 역할이 점차 증대되고 있으며, 향후 이러한 추세가 지속될 것이라는 것을 보여준다. 한편, 일관적인 인구와 통행의 변화 방향에도 불구하고, 철도와 고속도로의 네트워크 변화 방향은 상반되게 나타났다. 고속도로 통행 네트워크는 대도시를 중심으로 인접한 도시들과의 연결성을 강화하는 방향으로 변화한 반면, 철도 통행 네트워크는 대도시 간 연결성을 강화하는 방향으로 변화하였다. 본 학위논문은 또한 도시간 교통 수요를 예측할 수 있는 모델을 제시한다. 도시간 통행간의 공간적, 시간적 상관성 및 수요의 큰 이질성을 고려하기 위해 계층화된 그래프 합성곱 신경망 모델을 제시하였다. 본 연구는 그래프 이론, 딥러닝, 교통공학을 접목하여 도시간 인간 이동의 다양한 특성을 규명하고 모델링 하였으며, 그동안 풀지 못했던 교통 문제를 다학제적 연구를 통해 풀었다는 점에서 의의가 있다.

서지기타정보

서지기타정보
청구기호 {DGT 20005
형태사항 iv, 110 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 여지호
지도교수의 영문표기 : Kitae Jang
지도교수의 한글표기 : 장기태
학위논문 학위논문(박사) - 한국과학기술원 : 조천식녹색교통대학원,
서지주기 References : p. 102-107
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

표/그림

모든 표/그림 보기

Research Organization

Studies estimating traffic flow and demand

The nodes for the inter-city travel network: nodes are defined as th administrative cities ofSouth Korea. There are 115 nodes for highway and 76 forrailway ou of161 nodes in 2016

The number of nodes of the highway and railway networks over the years

The total number of trips occurred by highway and railway over the years

The total length ofthe highway and car kilometer traveled ofthe railway: the length ofthehighway increased continuously, while the car-kilometer ofthe railway increased discretely

Cumulative distribution function (CDF) ofdegree of highway (a) and railway (b) networks from 1977 (red color) to 2016 (blue color): Degree distributions have the homogeneous pattern in both networks, which is far from the scale-free property

The number of edges in the inter-city travel network of the railway: In the 1970s and 1980s, the number oflinks dropped sharply

Complementary Cumulative distribution function (CCDF) ofstrength (S) of the highway and railway networks from 1977 (red color) to 2016 (blue color): As time changes, both networks have evolved in the direction of intensifying heavy-tail tendency having straightlines at tails

Parameters for power-law distribution and goodness-of-fittests ofthe strength distribution (highway): Xmin is lower bound of power-law distribution; ac is power-law coefficient; p-value ofgofis the result ofthe goodness-of-fit test using a bootstrap procedure. The higher p-value indicates power-law is plausible; Log-likelihood ratio (LR) and high p- value ofLR reveal thatlog-normal distribution do

Parameters for power-law distribution and goodness-of-fittests ofthe strength distribution (railway)

Changes ofestimated power-law coefficients overtheyears ofthe highway and railway travel networks: both networks have evolved in the direction ofintensifying heavy- tail tendency

Lorenz curve of node strength ofhighway (a) and railway network (b)

Gini coefficient ofnode strength of highway (a) and railway network (b)

(a) Complementary Cumulative Density Function (CCDF of population distribution and (b) changes of power-law coefficients over the years: Population distribution also follows power-law, and the heavy-tail tendency has continuously strengthened over the years

Population distribution in a spatial region: As time went by, medium-sized cities (from orange to purple) shrank, and population polarization has intensified with a clear distinction between extremely small (yellow) andlarge cities (navy) 47

Travel demand ofhighway travels versus population in cities in South Korea from 1980 to 2015: Red lines show the bestfi toascalingrelation Y(N) = YoNB Scalingexponents arefurnished witha95%confidence interval. Scalingexponentsofinter-city travel wer sublinear (smaller than 1) but steadily increased over the years. It represents that the peopleliving in a largecity have increasingly stronge connect

Travel demand ofrailway travels versus population in cities in South Korea from 1980 to 2015: Scalingexponents of inter- city travel ofrailway were also sublinear, butthe relationship has changed from sublinear to linearin the railway

Complementary Cumulative distribution function (CCDF) ofweights (w) of the highway and railway networks from 1977 (red color) to 2016 (blue color): Weight distributions decay faster attail showing the log-normal distribution rather than the power-law. Itis different from the weights ofthe aviation network, which follows power-law.

Parameters for a log-normal distribution and goodness-of-fit testofthe weight distribution (highway): Thenegativesign oflog-likelihood ratio and low p-valueindicate that the log-normal distribution is more suitable for the observations than the power-law distribution

Parameters for a log-normal distribution and goodness-of-fittest ofthe weight distribution (railway)

Density plots of weight distribution in log-scaled

Changing ofinter-city travel network of the highway in the spatial region: Connections among neighboring cities become stronger. Especially, stronger connectivitybetween theregional hubs andtheiradiacent cities.

Changing ofinter-city travel network ofthe railway in thespatialregion: The railway travel network has evolved intensifvin, the connection betweendistanthub cities

Weighted rich-club effect of the highway and railway networks: In the highway network, there is a strong rich-club tendency over 80 - 95 percentile ofthe nodes. In the railway network, on the contrary, the magnitude of the rich-club effect has the highest values among thetop 5 - 10% ofnodes, which indicates that therailway network has evolved in a way to strengthen travels among thetop 5% hub citi

Weighted rich-club effect of the railway network before and after the introduction ofthe high-speed railway system: Rich club effects had risen significantly after 2004when the high-speed railway was introduced.

Community structureofthe highway travel network: Thecommunity ofthehighwaytravelnetworkhasevolvedformingstrong regional clusters. It1S very similar to the metropolitan economic zones specified by the National Territory ComprehensivePlanestablished by the government every 20 years

Community structure of the railway travel network: In the 1970s and 1980s, communities had formed regionally showing spatial constraints. After the 2000s, however,the community has formed alongthe railway route. In 2016, Seoul (northwestward) and Busan (southeastward), the two farthestand biggestmetropolitan cities in South Korea are clustered as one community.

The correlation ofthe first differences between traffic demands among major cities in South Korea from 2015 to 2018. There are three types of spatial dependencies between O-D traffic demands: 1) 0-D pairs having the same nodes but opposite direction, 2) 0-D pairshavingthe same origin or destination, and 3) 0-D pairsfrom and to similar regions.

Illustration of the graph structure ofO-D pairs

Daily Origin-Destination traffic demand distribution in 2019: (a) Histogram of the demand and (b) Complementary Cumulative Density Function (CCDF) of demand. The 7ㄷ

Partitions for the stratified model. The whole O-D graphis divided into foursub- graphs by the mean value of each O-D pair over time E(qk). Pn are the critical values for dividing the partitions. They are set 1, 100, and 10,000, respectively.

Illustration of the concept of the stratified GCN framework

Description of time embedding

The structure ofthe GCN module: (a) GCN block thatapplies skip connection and (b) GCN module that stacks multiple GCN blocks and fuses the time embedding vector at the end of GCN block.

Actual versus predicted demand of the stratified GCN (a and b) and the unified CNN (cand d) ofpartition 1 and 2: The stratified GCN outperforms the unified CNN in both evaluation metrics, Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). However, the stratified GCN tends to underestimate the demands nearthe upper boundary ofeach partition (See the dotted circle in Figures 4

The unified CNN model: The model predicts the whole O-D demand matrix a once. The structure ofthe modelis similar to the stratified GCN module in Figure 4-4b excep that the CNN model stacks multiple 1x1 convolutional filters instead of grapl convolutional filters.

The outliers of the origin-destination travel demand data: The values are considered outliers ifthey exceed 3.29 times ofstandard deviation (o(·)) from the mean value (E(:)). The horizontal red lines are the upper and lower boundary ofthe outliers. The outliers are substituted for the values from the previous week.

All the hyper-parameters of the model: Bold means selected ones.

Performance of the model compared with different baselines

Visualization oftheprediction accuracy: (a) The total travel demand ofpredicted and actual demand over 180daysofthe test set and (b)Ascatter plot ofactual versus predicted demand ofthe test set.