New Jersey train delay prediction and on-time rate improvement

Sijia Zhang & Yixuan Zhou

2023-11-26

For Product Introduction, visit NJT - Connect is Ready!.

1. Use Case

Our initiatives aim to effectively address and minimize delays for NJ Transit, ultimately reducing the financial and time-related setbacks experienced by New Jersey Transit and its passengers.

In 2019, New Jersey Transit faced a higher number of outages compared to any other system in the country, as reported by the Federal Transit Administration. And the on-time performance rate this year has fallen below target goal.

Figure 1.1. NJ Transit Train

The reliability of New Jersey’s train services is paramount for commuters and travelers, with disruptions potentially leading to legal consequences, damages, and increased operating costs. These challenges are exacerbated by negative word of mouth, diminished ridership, and the risk of losing market share, especially amid fierce competition with Amtrak.

Our app, NJ Connect, empowers NJ Transit staff by providing advance notice of potential delays and enabling proactive measures to mitigate or avoid disruptions. NJ Connect comes in two versions: a computer version for the command room and a mobile version for all station staff.

Figure 1.2 Web Interface Workflow

The computer version for the command room enables real-time monitoring and prediction of delays, as well as the management of trains and staff. Potential management measures include:

1. Train Management

• Schedule Adjustments: Introduce shuttle buses during peak passenger flow periods or adjust train speeds and dwell times.
• Priority Management: Assign priority status to certain trains or routes to minimize the impact of delays on critical services or high-capacity routes.
• Rerouting: Rearrange spare tracks based on the time needed for track fault repairs.

1. Staff Management
   • Issue Instructions
   • Receive Reports
2. Passenger Communication
   • Announcements
   • Digital Displays

Manage real-time information displayed on digital screens regarding train schedules, delays, and platform changes.

Figure 1.3 Delay Prediction for NJ Transit (Web Interface)

Figure 1.4 Delay Prediction for NJ Transit (Web Interface)

The mobile version for staff enables NJ Transit employees to:

1. View real-time and predicted delays.

2. Receive commands from the command center and indicate their execution or completion with a simple click.

3. Report to the command center any issues that could cause delays.

   

2. Data Wrangling

In this section, we present the data sources involved in this report and the initial data cleaning and integration methods used to prepare for subsequent statistical analysis and model construction.

2.1 NJ Station Data

To depict the spatial location of the New Jersey Transit rail network in New Jersey and surrounding areas, we used data on the locations of New Jersey Transit stations and lines. Concurrently, we incorporated New Jersey’s county-level geographical areas to present the New Jersey Transit rail network with 14 rail lines as spatial pattern.

Show

Figure 2.1. New Jersey Transit Lines and Stations Map

2.2 Delays by station and line

Delays invariably occur in accordance with the unique features of each station and every railway line. In this section, we use a dataset from Kaggle that list origin/destinations delay for Amtrak and NJ Transit trains. This dataset contains monthly records that detail the performance information for nearly all train trips across the NJ Transit rail network. We extract the rail data from October to November 2019, which included valuable information on train information, line, and delay time.

Show

2.3 Weather Data

Next, we explore the weather data pulled from the RIEM package in R, including four features: temperature (Fahrenheit), precipitation in the past hour (Inches), visibility (Miles), and wind speed (Knots). The complex variations in weather can impact decision-making regarding railway transit networks. Generally, high temperatures, substantial precipitation, and low visibility do not severely affect the normal operation of rail systems. However, extreme, and atypical weather conditions such as typhoons, blizzards, heavy downpours, and hailstorms can be decisive factors in preventing trains from operating normally on tracks. Therefore, we incorporated weather data as a reference data in our further predictive models.

Show

Figure 2.2 Weather data - Newark Liberty International Airport (EWR) - Oct & Nov 2019

2.5 Create Space-Time Panel and Time Lag Features

In this report, to better understand the delays in New Jersey’s transit railways, we create a time series from a spatio-temporal perspective to analyze the train delays. We hypothesize that when a delay occurs at a station on a railway line, it causes subsequent delays at other stations on the line. Therefore, we establish predictive models by forecasting the delays at each station on every line for each hour interval and every daily interval during the study period to predict the spatial-temporal process of delays.

Show

Show

Figure 2.3. Mean Delays of Each Lines

3. Exploratory Analysis

Before developing models, we explored the delay data over time, across space, across time and space, as well as the influences by time lag and weather.

3.1 On-time performance overview

We began by examining the overview of delay frequency. As illustrated in the graph below, the frequency of delays diminishes with the passage of time. Notably, there is a scarcity of trains experiencing delays exceeding 30 minutes.

Show

Figure 3.1 Overview of NJ Transit Delays at All Stations

To provide a quantitative assessment, we computed the percentage of trains adhering to their scheduled arrival times. Merely 68% of trains managed to arrive within five minutes of the expected time, and only 89% achieved this within a ten-minute window. When extending the timeframe to 20 minutes, a substantial improvement is observed, with almost all (97%) trains arriving punctually. Remarkably, all trains successfully reach their destinations within a two-hour timeframe. This analysis underscores the severity of the train delay issue, posing potential setbacks for both NJ Transit and its patrons.

Show

Time_Frame	Percentage_of_trains
Arrived within 5 minutes	68.00
Arrived within 10 minutes	88.67
Arrived within 20 minutes	97.07
Arrived within 30 minutes	98.68
Arrived within 1 hour	99.64
Arrived within 2 hours	100.00

The plot also reveals an alarming trend—nearly all on-time rates for trains fall below 0.5, with over half plummeting below 0.1. This implies a disconcerting scenario where the likelihood of trains arriving on schedule is exceedingly rare.

Show

Figure 3.2 On-Time rate of Trains by NJ Transit Stations

Moving on to the on-time rate maps for trains, a geographical pattern emerges. Stations in Gloucester and Atlantic exhibit markedly lower on-time rates compared to those in the northern regions of New Jersey. This geographic disparity in on-time performance raises concerns about the consistency of service across different areas.

Show

Figure 3.3 On-Time rate of Trains by NJ Transit Stations within 5/10/20/30 minutes

3.2 Delays across time

Secondly, we delved into the temporal dimension, scrutinizing delays on both a weekly and daily basis.

When examining the total train stops and average delay minutes over a week, a discernible cyclical pattern emerges. This cycle, however, becomes more intricate when analyzed on a daily scale.

Show

Figure 3.4 Average Train Delays of Each Line

Upon closer inspection of day-to-day variations, we observe that the average delay time exhibits a remarkably consistent fluctuation pattern from Monday to Friday. The shortest delay occurs at 3:30 in the morning, gradually increasing until it peaks around 9:00. Subsequently, there is a modest decrease until 15:00, followed by a continuous extension until 19:00 when it reaches the day’s highest peak.

During the weekends, the overall pattern of delay time fluctuations mirrors that of weekdays. However, the timing of the lowest and highest peaks differs, occurring later than the corresponding times on weekdays. Notably, both the overall average delay time and the peak value during weekends surpass those recorded from Monday to Friday.

Show

Figure 3.5 Average Train Delays per hr. by Day

3.3 Delays across space

Thirdly, we start to explore the delays across space (by lines and by stations).

(1) Delays by lines

Examining the table below reveals notable variations in the average delay durations across different transit lines. The most considerable delay is observed on the Atl. City Line, reaching up to 17 minutes. Notably, the Princeton Shuttle Line stands out for its minimal delay, registering only 0.4 minutes, indicating a nearly negligible delay scenario.

Show

Table 3.3.1 Summary of Delays by Line

line	Mean_Delay_Minutes	Mean_Delay_Hours	Median_Delay_Minutes	Median_Delay_Hours	Min_Delay_Minutes	Max_Delay_Minutes
Atl. City Line	16.5593655	0.2759894	0.0750000	0.0012500	0	121.08333
Bergen Co. Line	4.3744556	0.0729076	3.0833333	0.0513889	0	99.00000
Gladstone Branch	4.4351968	0.0739199	2.7166667	0.0452778	0	100.21667
Main Line	4.4068482	0.0734475	3.1833333	0.0530556	0	100.00000
Montclair-Boonton	4.9462978	0.0824383	3.2166667	0.0536111	0	90.00000
Morristown Line	5.4243676	0.0904061	3.4333333	0.0572222	0	100.50000
No Jersey Coast	4.7192779	0.0786546	3.1500000	0.0525000	0	91.43333
Northeast Corrdr	3.8530196	0.0642170	2.2000000	0.0366667	0	69.33333
Pascack Valley	3.4937946	0.0582299	2.4166667	0.0402778	0	98.00000
Princeton Shuttle	0.3816861	0.0063614	0.0166667	0.0002778	0	11.46667
Raritan Valley	4.3672723	0.0727879	3.0666667	0.0511111	0	81.00000

To enhance comprehension, we translated these findings into visual representations using bar plots and a map. These visualizations offer a clear depiction of the severity of delays on each line and their respective geographic locations.

Show

Figure 3.6 Average Delay by Lines in NJ

(2) Delays by stations

Given that the average delays across lines fail to pinpoint the exact locations of specific delay incidents, we further scrutinize the delays categorized by stations. As illustrated in the table 3.3 below, the majority of stations experience delays of approximately 5 minutes. However, it is noteworthy that 8 stations exhibit mean delay durations surpassing 15 minutes, with the most prolonged delay reaching 18 minutes at Lindenwold.

Show

Table 3.3.2 Summary of Delays by Stations

from	Mean_Delay_Minutes	Mean_Delay_Hours	Median_Delay_Minutes	Median_Delay_Hours	Min_Delay_Minutes	Max_Delay_Minutes
Aberdeen-Matawan	5.4631657	0.0910528	4.0333333	0.0672222	0.0000000	67.31667
Absecon	16.4318971	0.2738650	0.0000000	0.0000000	0.0000000	116.00000
Allendale	4.2223364	0.0703723	3.0500000	0.0508333	0.0000000	98.00000
Allenhurst	5.4804540	0.0913409	4.2000000	0.0700000	0.0000000	50.00000
Anderson Street	4.2098418	0.0701640	3.2250000	0.0537500	0.0000000	67.00000
Annandale	3.5718182	0.0595303	1.4916667	0.0248611	0.0000000	50.00000
Asbury Park	5.1515504	0.0858592	4.1000000	0.0683333	0.0000000	49.00000
Atco	16.8217885	0.2803631	0.1333333	0.0022222	0.0000000	118.05000
Atlantic City Rail Terminal	16.7359309	0.2789322	8.0166667	0.1336111	0.0000000	85.38333
Avenel	5.1985250	0.0866421	4.0500000	0.0675000	0.0000000	64.00000
Basking Ridge	4.9240414	0.0820674	3.0500000	0.0508333	0.0000000	98.00000
Bay Head	1.8547227	0.0309120	1.2083333	0.0201389	0.0000000	40.15000
Bay Street	5.1028701	0.0850478	3.4333333	0.0572222	0.0000000	88.00000
Belmar	6.2048588	0.1034143	5.0833333	0.0847222	0.0000000	49.00000
Berkeley Heights	5.7971439	0.0966191	4.1666667	0.0694444	0.0000000	98.00000
Bernardsville	5.5911142	0.0931852	4.0833333	0.0680556	0.0000000	100.21667
Bloomfield	4.6073370	0.0767889	3.2000000	0.0533333	0.0000000	71.86667
Boonton	5.8770526	0.0979509	3.4666667	0.0577778	0.0000000	69.00000
Bound Brook	4.8422389	0.0807040	3.2333333	0.0538889	0.0000000	78.00000
Bradley Beach	5.5062292	0.0917705	4.2166667	0.0702778	0.0000000	49.00000
Brick Church	5.6478944	0.0941316	4.1000000	0.0683333	0.0000000	98.00000
Bridgewater	5.2014232	0.0866904	3.4666667	0.0577778	0.0000000	81.00000
Broadway Fair Lawn	3.8357583	0.0639293	2.4333333	0.0405556	0.0000000	53.11667
Campbell Hall	7.5154343	0.1252572	6.1500000	0.1025000	0.0000000	99.00000
Chatham	5.6988135	0.0949802	3.9916667	0.0665278	0.0000000	98.00000
Cherry Hill	18.0412492	0.3006875	2.8666667	0.0477778	0.0000000	121.08333
Clifton	5.0168803	0.0836147	4.2000000	0.0700000	0.0000000	65.05000
Convent Station	5.8465857	0.0974431	4.1333333	0.0688889	0.0000000	99.00000
Cranford	4.5691091	0.0761518	3.1666667	0.0527778	0.0000000	77.00000
Delawanna	4.6027406	0.0767123	3.6583333	0.0609722	0.0000000	63.33333
Denville	5.0666106	0.0844435	3.0166667	0.0502778	0.0000000	99.00000
Dover	3.3204245	0.0553404	2.0166667	0.0336111	0.0000000	71.45000
Dunellen	4.8293326	0.0804889	3.2500000	0.0541667	0.0000000	78.00000
East Orange	5.4308598	0.0905143	3.6916667	0.0615278	0.0000000	98.00000
Edison	3.9794028	0.0663234	2.1833333	0.0363889	0.0000000	49.35000
Egg Harbor City	17.5627215	0.2927120	1.1333333	0.0188889	0.0000000	119.16667
Elberon	4.3321162	0.0722019	3.0500000	0.0508333	0.0000000	50.00000
Elizabeth	4.7750499	0.0795842	3.3000000	0.0550000	0.0000000	65.00000
Emerson	3.7367694	0.0622795	3.0666667	0.0511111	0.0000000	69.00000
Essex Street	4.0836080	0.0680601	3.1500000	0.0525000	0.0000000	98.00000
Fanwood	4.8693558	0.0811559	3.3500000	0.0558333	0.0000000	77.00000
Far Hills	5.4844938	0.0914082	3.5583333	0.0593056	0.0000000	85.00000
Garfield	4.6685518	0.0778092	3.3166667	0.0552778	0.0000000	55.05000
Garwood	5.3126951	0.0885449	3.7416667	0.0623611	0.0000000	77.00000
Gillette	4.7553597	0.0792560	2.8583333	0.0476389	0.0000000	98.00000
Gladstone	0.8500000	0.0141667	0.4500000	0.0075000	0.0000000	12.15000
Glen Ridge	5.1392524	0.0856542	3.5166667	0.0586111	0.0000000	88.00000
Glen Rock Boro Hall	4.4068755	0.0734479	3.2833333	0.0547222	0.0000000	50.23333
Glen Rock Main Line	4.1949211	0.0699154	3.2333333	0.0538889	0.0000000	70.03333
Hackettstown	1.2590847	0.0209847	0.4250000	0.0070833	0.0000000	16.13333
Hamilton	2.1174674	0.0352911	0.5500000	0.0091667	0.0000000	58.00000
Hammonton	16.2168675	0.2702811	0.0750000	0.0012500	0.0000000	108.63333
Harriman	7.8354281	0.1305905	6.2583333	0.1043056	0.0000000	98.00000
Hawthorne	4.4457846	0.0740964	3.5166667	0.0586111	0.0000000	70.15000
Hazlet	5.1892691	0.0864878	3.3833333	0.0563889	0.0000000	68.13333
High Bridge	0.7634021	0.0127234	0.0000000	0.0000000	0.0000000	31.28333
Highland Avenue	5.5215899	0.0920265	3.7500000	0.0625000	0.0000000	98.00000
Hillsdale	3.8776495	0.0646275	3.1166667	0.0519444	0.0000000	68.00000
Hoboken	2.8702549	0.0478376	2.1000000	0.0350000	0.0000000	91.43333
Jersey Avenue	3.6976969	0.0616283	2.1500000	0.0358333	0.0000000	51.00000
Kingsland	4.3889168	0.0731486	3.3000000	0.0550000	0.0000000	61.06667
Lake Hopatcong	4.3083333	0.0718056	3.1083333	0.0518056	0.0000000	45.00000
Lebanon	5.0425038	0.0840417	2.1833333	0.0363889	0.0000000	57.00000
Lincoln Park	6.5301404	0.1088357	4.4333333	0.0738889	0.0000000	69.00000
Linden	4.7829857	0.0797164	3.2833333	0.0547222	0.0000000	65.00000
Lindenwold	17.9770103	0.2996168	2.1500000	0.0358333	0.0000000	120.08333
Little Falls	6.4842105	0.1080702	5.0666667	0.0844444	0.0833333	68.00000
Little Silver	3.7825894	0.0630432	2.3833333	0.0397222	0.0000000	85.00000
Long Branch	2.0822455	0.0347041	0.2666667	0.0044444	0.0000000	70.13333
Lyndhurst	4.5456708	0.0757612	3.6000000	0.0600000	0.0000000	63.13333
Lyons	6.2031404	0.1033857	4.1500000	0.0691667	0.0000000	98.00000
Madison	5.7825121	0.0963752	4.0833333	0.0680556	0.0000000	99.00000
Mahwah	3.8536238	0.0642271	2.3333333	0.0388889	0.0000000	98.00000
Manasquan	4.4067414	0.0734457	2.2333333	0.0372222	0.0000000	49.00000
Maplewood	5.4635546	0.0910592	4.0666667	0.0677778	0.0000000	100.41667
Metropark	4.1932944	0.0698882	3.0500000	0.0508333	0.0000000	56.35000
Metuchen	4.0716912	0.0678615	2.3916667	0.0398611	0.0000000	49.00000
Middletown NJ	5.4959119	0.0915985	4.0166667	0.0669444	0.0000000	84.00000
Middletown NY	6.6444242	0.1107404	4.5666667	0.0761111	0.0000000	98.00000
Millburn	5.3709676	0.0895161	3.5750000	0.0595833	0.0000000	98.00000
Millington	4.9956363	0.0832606	3.1500000	0.0525000	0.0000000	99.00000
Monmouth Park	9.6904167	0.1615069	9.0000000	0.1500000	0.0000000	31.01667
Montclair Heights	4.8627932	0.0810466	3.1000000	0.0516667	0.0000000	90.00000
Montclair State U	2.5546321	0.0425772	1.2583333	0.0209722	0.0000000	71.10000
Montvale	3.5182448	0.0586374	2.3000000	0.0383333	0.0000000	68.00000
Morris Plains	5.2196153	0.0869936	3.2666667	0.0544444	0.0000000	99.08333
Morristown	5.7324428	0.0955407	4.1000000	0.0683333	0.0000000	100.00000
Mount Arlington	5.8279595	0.0971327	4.1500000	0.0691667	0.0000000	53.00000
Mount Olive	3.6589620	0.0609827	2.2000000	0.0366667	0.0000000	45.00000
Mount Tabor	4.3404744	0.0723412	2.9500000	0.0491667	0.0000000	84.00000
Mountain Avenue	5.8009993	0.0966833	4.0166667	0.0669444	0.0000000	88.00000
Mountain Lakes	5.5762807	0.0929380	3.3333333	0.0555556	0.0000000	69.00000
Mountain Station	5.4338421	0.0905640	3.4333333	0.0572222	0.0000000	100.03333
Mountain View	6.5078246	0.1084637	5.0500000	0.0841667	0.0500000	69.00000
Murray Hill	3.5357428	0.0589290	1.4166667	0.0236111	0.0000000	98.00000
Nanuet	3.2924045	0.0548734	2.2166667	0.0369444	0.0000000	73.00000
Netcong	5.2204019	0.0870067	3.4166667	0.0569444	0.0000000	45.00000
Netherwood	4.5553783	0.0759230	3.1666667	0.0527778	0.0000000	77.00000
New Bridge Landing	3.4179578	0.0569660	2.4500000	0.0408333	0.0000000	67.00000
New Brunswick	3.2671977	0.0544533	1.9500000	0.0325000	0.0000000	57.00000
New Providence	4.8228021	0.0803800	3.1166667	0.0519444	0.0000000	98.00000
New York Penn Station	3.1828261	0.0530471	1.1500000	0.0191667	0.0000000	90.13333
Newark Airport	4.5656323	0.0760939	3.1833333	0.0530556	0.0000000	69.33333
Newark Broad Street	5.4070034	0.0901167	3.5166667	0.0586111	0.0000000	99.00000
Newark Penn Station	5.0741233	0.0845687	3.4166667	0.0569444	0.0000000	76.00000
North Branch	6.7292906	0.1121548	4.1166667	0.0686111	0.0000000	58.00000
North Elizabeth	4.8466839	0.0807781	3.3833333	0.0563889	0.0000000	65.00000
Oradell	3.3879580	0.0564660	2.3500000	0.0391667	0.0000000	68.00000
Orange	5.5007296	0.0916788	4.0333333	0.0672222	0.0000000	98.00000
Otisville	6.4622680	0.1077045	4.1166667	0.0686111	0.0000000	100.00000
Park Ridge	3.9080446	0.0651341	3.1166667	0.0519444	0.0000000	68.00000
Passaic	4.4787440	0.0746457	3.5583333	0.0593056	0.0000000	64.23333
Paterson	4.5937419	0.0765624	4.0666667	0.0677778	0.0000000	64.10000
Peapack	4.9507037	0.0825117	2.1333333	0.0355556	0.0000000	99.00000
Pearl River	3.5678114	0.0594635	2.2000000	0.0366667	0.0000000	75.00000
Pennsauken	17.7121026	0.2952017	2.1333333	0.0355556	0.0000000	121.00000
Perth Amboy	5.2422320	0.0873705	3.5000000	0.0583333	0.0000000	67.23333
Philadelphia	11.3932609	0.1898877	0.0000000	0.0000000	0.0000000	109.38333
Plainfield	4.7577553	0.0792959	3.2500000	0.0541667	0.0000000	78.00000
Plauderville	3.7120678	0.0618678	2.2833333	0.0380556	0.0000000	54.13333
Point Pleasant Beach	2.9806799	0.0496780	1.2500000	0.0208333	0.0000000	50.00000
Port Jervis	1.8889984	0.0314833	0.0000000	0.0000000	0.0000000	77.26667
Princeton	0.3820283	0.0063671	0.0166667	0.0002778	0.0000000	11.46667
Princeton Junction	3.4141258	0.0569021	1.4833333	0.0247222	0.0000000	55.00000
Radburn Fair Lawn	4.5040875	0.0750681	3.2333333	0.0538889	0.0000000	98.00000
Rahway	4.9242468	0.0820708	3.2666667	0.0544444	0.0000000	66.00000
Ramsey Main St	4.4770522	0.0746175	3.1166667	0.0519444	0.0000000	98.00000
Ramsey Route 17	4.8100658	0.0801678	3.1500000	0.0525000	0.0000000	98.00000
Raritan	3.4545129	0.0575752	2.2000000	0.0366667	0.0000000	56.00000
Red Bank	5.6495212	0.0941587	4.0500000	0.0675000	0.0000000	85.00000
Ridgewood	4.5506652	0.0758444	3.2916667	0.0548611	0.0000000	98.00000
River Edge	3.3353687	0.0555895	2.2833333	0.0380556	0.0000000	68.00000
Roselle Park	4.7237920	0.0787299	3.2000000	0.0533333	0.0000000	76.00000
Rutherford	5.1104014	0.0851734	4.0666667	0.0677778	0.0000000	98.00000
Salisbury Mills-Cornwall	7.8822236	0.1313704	6.2166667	0.1036111	0.0000000	98.00000
Secaucus Concourse	0.1000000	0.0016667	0.1000000	0.0016667	0.1000000	0.10000
Secaucus Lower Lvl	3.7083083	0.0618051	3.0000000	0.0500000	0.0000000	99.00000
Secaucus Upper Lvl	4.2567090	0.0709452	2.3500000	0.0391667	0.0000000	79.00000
Short Hills	5.5491160	0.0924853	3.4666667	0.0577778	0.0000000	98.00000
Sloatsburg	6.8826422	0.1147107	5.2166667	0.0869444	0.0000000	98.00000
Somerville	3.8435295	0.0640588	2.3666667	0.0394444	0.0000000	80.00000
South Amboy	5.5509681	0.0925161	4.0666667	0.0677778	0.0000000	67.11667
South Orange	5.8126665	0.0968778	4.1166667	0.0686111	0.0000000	99.06667
Spring Lake	5.3182724	0.0886379	4.1333333	0.0688889	0.0000000	49.00000
Spring Valley	1.6377672	0.0272961	1.2000000	0.0200000	0.0000000	56.30000
Stirling	4.5160907	0.0752682	2.6833333	0.0447222	0.0000000	98.00000
Suffern	3.4095299	0.0568255	1.4666667	0.0244444	0.0000000	98.00000
Summit	5.1305265	0.0855088	3.2833333	0.0547222	0.0000000	100.50000
Teterboro	4.0396877	0.0673281	3.2000000	0.0533333	0.0000000	42.56667
Towaco	5.5052982	0.0917550	3.3166667	0.0552778	0.0000000	69.00000
Trenton	2.1226712	0.0353779	1.1833333	0.0197222	0.0000000	43.05000
Tuxedo	7.1140173	0.1185670	5.3500000	0.0891667	0.0000000	98.00000
Union	4.2109883	0.0701831	3.0000000	0.0500000	0.0000000	79.00000
Upper Montclair	5.8551133	0.0975852	4.1500000	0.0691667	0.0000000	89.00000
Waldwick	3.2202804	0.0536713	2.1666667	0.0361111	0.0000000	69.00000
Walnut Street	5.3643026	0.0894050	3.5666667	0.0594444	0.0000000	88.00000
Watchung Avenue	5.6386605	0.0939777	4.1166667	0.0686111	0.0000000	88.00000
Watsessing Avenue	5.5358640	0.0922644	4.2083333	0.0701389	0.0000000	76.00000
Wayne-Route 23	6.8869474	0.1147825	5.1333333	0.0855556	0.1000000	68.00000
Wesmont	4.3013032	0.0716884	3.1666667	0.0527778	0.0000000	56.15000
Westfield	4.5479696	0.0757995	3.1416667	0.0523611	0.0000000	77.00000
Westwood	4.0529744	0.0675496	3.2000000	0.0533333	0.0000000	69.00000
White House	4.5695205	0.0761587	1.3416667	0.0223611	0.0000000	56.00000
Wood Ridge	4.5781898	0.0763032	3.6000000	0.0600000	0.0000000	66.00000
Woodbridge	4.9260991	0.0821017	3.2000000	0.0533333	0.0000000	53.06667
Woodcliff Lake	4.0913741	0.0681896	3.2000000	0.0533333	0.0000000	47.30000

Figure 3.7 Average Delay by Stations in NJ

3.4 Delays across time and space

Fourthly, we investigate temporal (weekly and daily) and spatial delays.

Presented below are the daily mean delays throughout the week, outlined by day. As depicted in the plot, the average delays in the northern region of New Jersey range consistently below 10 minutes from Monday to Sunday. In contrast, the mean delay for the AC line exceeds 10 minutes, particularly on weekends where it peaks at 20 minutes.

Show

Figure 3.8 Mean Delay Minutes per Day by Line

Additionally, we examine the hourly delays categorized by line.

Show

Figure 3.9 Mean delay minutes per hour by line

Show

3.6 Delays impacted by weather

In this section, we investigate the lack of any significant correlation between train delays and the various weather conditions. And the reason for the result might be that only severe weather conditions are likely to lead to significant train delay. Our model was trained on a limited time period, and no noteworthy weather events occurred during that period to significantly manifest in the data. Although their impact was only marginal, our final model incorporated temperature, precipitation, visibility, and wind speed features due to their slight predictive enhancements.

Show

Figure 3.9 Train delays by weather conditions

4. Regression model building and training

In this part, we began to divide train and test set, build 5 different models.

4.1 Create Train and Test Set

We have the delay data from the 40th week to the 46th week. Considering that we calculated the lag data, there are the complete data from 41th week to the 46th week. Then we filter the first four weeks as our train set, and the last two weeks as our test set.

Show

Figure 3.10 NJ Transit Rail Trips by Week

4.2 Correlation Matrix

The correlation matrix shows the relationship across all the numeric variables. From Figure 3.10 below, there is no obvious correlation between any two variables. Thus, we kept all the variables for the models we built later.

Show

Figure 3.11 Correlation Matrix

4.3 Build Model

During the development of our model, we categorized the features we incorporated into four broad groups. These categories were systematically tested in various combinations, leading us to the presentation of five models, among which the one we ultimately selected is featured.

Here are our four feature categories:

• Time: week (Week number), dotw (Weekday), hour (Time of a day)
• Space: ST_NAME (Name of stations), LINE_CODE (Abbreviation for lines)
• Time lag: lag1Hour, lag2Hours, lag3Hours,lag4Hours, lag6Hours, lag8Hours, lag10Hours, lag12Hours
• Weather: temperature, precipitation, visibility, windspeed
• Holidays:holidayLag

Here are our three models:

• 1. Time and Space
• 1. Time, Space and Time Lag
• 1. Time, Space, Time Lag and Weather
• 1. Time, Space, TIme Lag, Weather and HolidayLag

5. Results

5.1 Modeling for Training Set

5.2 Predict for Test Set

5.3 MAEs

We validate the models based on the Mean Absolute Error (MAE) over time and can see that model B and Model C has the smallest MAE, whether it is calculated by week and model or by line.

Show

Figure 5.1 MAE: Delays by week and model

Show

Figure 5.2 MAE: Delays of more than 20 minutes by model and lines

5.4 Observed and Predicted Mean Hourly Delays

From predictions of two weeks’ data, we can see a significant enhancement in the ability of models B, C and D to predict highest peaks of delay. Also, they accurately predict low delay peaks and trends, meaning that incorporating time lags improves prediction accuracy. However, the models fail to precisely predict timing of high delay. We think it could be due to unforeseen situations.

Show

Figure 5.3 Predicted and Observed mean delay minutes by hour

5.5 Cross Validation

We conduct daily cross-validation on our linear regression model. Each day of the week sequentially becomes part of the training set, while the other days act as the test set. Results show that the average Mean Absolute Error (MAE) stands at 0.493 minutes, with the maximum daily MAE reaching about 1.2. The bulk of the MAE values cluster between 0.2 and 0.6, indicating that our model displays a relatively consistent and low prediction error across test set. This consistency suggests that our model has commendable predictive accuracy. However, despite the average error being relatively low, there remains room for improvement, especially in reducing the maximum error values.

Show

Figure 5.4 LOGO-CV: Mean Delay Duration by Day in a Week

Show

Table 5. Cross Validation Results

CrossValidation	Mean_MAE	SD_MAE
LOGO CV: day in a week	0.493	0.39

6. Conclusion

The model trained for the App reveals the primary factors leading to train delays, providing a multidimensional periodic delay prediction for NJ Transit that considers both temporal and spatial dimensions. This enables our model to proactively forecast delays by analyzing New Jersey’s traffic hub dynamics.

The utility and efficacy of our model are evident, yet its refinement depends on an extensive, precise, and publicly accessible database from governmental and specialized sectors. By reviewing literature and historical delay instances, we aim to mitigate unpredictable factors such as mechanical failures, funding shortages, and unexpected events. In 2023, for instance, the Morris & Essex line saw significant commuter disruptions due to ongoing overhead line issues. Such disruptions, especially from signal problems in critical areas like tunnels, can lead to notable delays due to safety and operational concerns.

To enhance our model, we need to integrate potential factors that may cause train delays or cancellations as independent variables. Another key to improving our delay prediction model’s accuracy is to include comprehensive Amtrak data, such as actual and estimated arrival times and train IDs, which enrich our independent variable observations.

Despite these considerations, we are confident that our model, which includes a wide array of important predictors, is robust enough for NJ Transit’s application. The train delay prediction model we’ve developed for NJ Transit aims to optimize operations, improve cost-effectiveness, and elevate the passenger experience and service quality. This data-driven approach not only assists NJ Transit in better understanding and managing its operational challenges but also supports its sustainable development goals over the long term. Moving forward, we are committed to continually enhancing our prediction model to offer even more valuable solutions to NJ Transit and other public transit operators.