← Back to selected research
Research Case Study · Urban Rail Reliability

Data-Driven Metro Travel-Time Reliability Assessment

I processed more than eight million Shanghai Metro smart-card records to study how travel-time reliability varies across origin–destination pairs. Our team combined reliability measures with anomaly detection and regression, then brought the results together in an operations dashboard.

InstitutionTongji University
PeriodOct 2024 – Jun 2025
Data8+ million Shanghai Metro AFC records
MethodsReliability metrics · Isolation Forest · Regression
>0.80Pairwise rank correlation among metrics
0.742Regression R²
0.00287Test-set MSE
Original research framework from the metro reliability report
Original project artifact · reliability distributions from the final report
Team project at Tongji University · My work focused on processing the AFC data, implementing and validating reliability indicators, and applying anomaly detection and regression.

Where and when are metro journeys less predictable?

Passengers making the same journey can experience different travel times because of demand, transfers, and operating conditions. We studied this variation for individual origin–destination (OD) pairs.

We first compared several reliability measures, then examined the spatial and temporal patterns of less reliable trips.

Analytical perspective

The measures describe travel-time dispersion, longer journeys in the tail of the distribution, and differences from a low-travel-time baseline.

From smart-card records to OD-level reliability.

Data cleaningExtract valid metro trips and construct OD-level travel-time samples.
Metric systemCompute CoV, TTI, BTI, and PTI to capture dispersion and long-tail behavior.
Consistency checkUse Spearman rank correlations to test whether the metrics identify similar reliability ordering.
Pattern miningIdentify low-reliability OD pairs and study external drivers using Isolation Forest, t-SNE, correlations, and regression.
Temporal & decision supportCompare periods and dates, then summarize findings in a manager-facing BI dashboard.

Four indicators describe different parts of the travel-time distribution.

MetricDefinition used in the projectMain interpretation
CoVCoefficient of variation of travel timeRelative dispersion
TTIMean travel time divided by free-flow travel time (15th percentile)Shift from low-end / free-flow travel time
BTIExtra time needed relative to the center of the distribution to arrive on time in most tripsLong-tail buffer
PTI95th-percentile travel time divided by 15th-percentile travel timeSpread between distribution tails

Spearman rank analysis showed that the four metrics were strongly consistent in how they ordered OD pairs: all reported pairwise correlations exceeded 0.80.

PairSpearman ρ
BTI – CoV0.8625
BTI – PTI0.9361
BTI – TTI0.8410
CoV – PTI0.9377
CoV – TTI0.9265
PTI – TTI0.9704
Original project figure
Distributions of four metro travel time reliability metrics
Figure 1. Distribution of the four reliability indicators across OD pairs. Original figure from the project report.
Original project figure
Chord diagrams showing OD reliability differences
Figure 2. OD-pair reliability chord diagrams. Variation in line intensity reflects substantial heterogeneity across the metro network.

Identifying OD pairs with low reliability.

After computing reliability values for all OD pairs, the project focused on the low-reliability tail. Isolation Forest was used to flag anomalous OD pairs, and t-SNE was used to visualize the resulting separation in a lower-dimensional space.

Original project figure
t-SNE visualization of anomalous metro OD pairs
Figure 3. t-SNE visualization of OD-pair anomalies. Red points denote the anomalous set identified for further investigation.

The study then examined whether unreliable OD pairs shared common attributes, including transfer involvement, passenger volume, central-city location, trip length / free-flow travel time, and station age.

External factors associated with reliability

External factorReported correlation with TTI
Free-flow travel time−0.8804
Passenger volume0.7130
Transfer involvement−0.2596
Central-city location0.3318
Origin station age0.1690
Destination station age0.2213

Free-flow travel time and passenger volume had the strongest reported correlations with TTI. We used them as predictors in the regression model.

Regression performance

The reported test-set mean squared error was 0.00287, with R² = 0.7421.

Reliability also changes by time of day and across dates.

The daily operating period was divided into five intervals: early-morning off-peak, morning peak, daytime off-peak, evening peak, and night off-peak. OD travel counts were used as weights when computing period-level mean PTI.

Original project figure
Weighted PTI by time period
Figure 4. Weighted PTI by operating period.
Original project figure
Daily weighted TTI trend
Figure 5. Daily weighted TTI trend.

In these data, the busiest periods were not always the least reliable. More regular service and passenger routines may help explain the pattern, although this comparison does not establish a causal relationship.

Bringing the results into an operations dashboard.

Our team built a dashboard so operators could inspect reliability across the network, compare operating periods, and locate OD pairs for further investigation.

Original research artifact · management dashboard
Original Shanghai Metro reliability dashboard
Figure 6. Original project dashboard built on Baidu Sugar BI.
Reading the dashboard

The panels show network reliability summaries, spatial patterns, and daily trends. They bring the results into one view so users can investigate an OD pair or operating period. The original dashboard uses Chinese interface labels.

I worked on OD-level data processing, reliability indicators, and pattern mining.

I processed more than eight million Shanghai Metro AFC records; implemented and validated CoV, TTI, BTI, and PTI; and applied Isolation Forest and regression modeling to identify unreliable OD pairs and examine associated factors.

I also helped visualize OD-level travel-time reliability so the results could support transit-management decisions.

Source: “数据驱动的地铁网络出行服务可靠性评估与规律挖掘,” Tongji University, final project report.

Conference paperCooperative Perception and Driving Behavior · TRB 2024 →Explore another projectIntelligent Train Driving Assistance →