Improving Bus Predictions Using Machine Learning

Ron Mok, Software Development Manager at TransLink

New Westminster, British Columbia, Canada

TransLink Quick Facts

• Regional transportation authority for Metro Vancouver

​

• Canada’s largest transit service area: 1800 km² (700 mi²)

​

• Serving a population of 2.5M residents

​

​

TransLink Quick Facts - Fleet

SkyTrain: 326 train cars that serve 53 stations along 79km (49mi) of rapid transit. Longest rapid transit system in Canada and the one of the longest fully automated driverless systems in the world.

​

SeaBus: 3 passenger-only ferries.

​

Bus: 200+ bus routes with a fleet of 1,500+ vehicles. 262 million bus trips across our network in 2018.

​

When Is The Next Bus?

Next Bus SMS

Next Bus Web App

Google Maps

Other 3^rd party apps

How Were Predictions Being Made?

• Predictions were based on comparing Automatic Vehicle Location (AVL) data to the schedule and shifting the schedule based on the difference.

​

• The AVL dataset provides real-time GPS data for buses. This data is sent whenever a bus leaves or passes by a bus stop or every 90 seconds whichever comes first.

​

• Susceptible to changing weather and traffic conditions as these predictions do not account for those.

​

• This is what we called our RTTI (Real Time Transit Information) system.

​

​

What Is The Problem?

• We have real-time next bus predictions, but they’re not always accurate.

​

What Is The Problem?

Our Approach – Machine Learning

Some Examples….

​

​

?

Run Time Model

Dwell Time Model

Additional Factors

?

B

A

Weather

• The Dark Sky API is used to acquire real-time and forecast weather data

​

• Transit system separated into 13 geographic-based weather regions

​

• Assumption: all segments & stops in a region experience the same weather

Weather

• The Dark Sky API offers a full collection of weather conditions. We used the following parameters in our models:

• Chance of precipitation
• Pressure
• Intensity of precipitation
• Wind speed
• Temperature
• Wind bearing

​

• Apparent temperature
• Cloud cover
• Dew point
• UV index
• Humidity
• Visibility

​

ML Details – Model Level Predictions

• The ML model we chose was Extreme Gradient Boosting (XGBoost) because of its fast training time and high-performance rate.

​

• Each bus stop and segment in the transit system has their own set of machine learning models, provided that there is enough data to train on (at least 500 instances of dwell/run events).

​

• As we create and train a different set of XGBoost models for every single segment and bus stop in our system, we end up with over 18,000 different sets of models.

​

​

ML Details – Bus Level Predictions

• With the ML models generating prediction times for individual segments and stops (model-level predictions), we need to create an algorithm that aggregates these predictions together to create stop arrival/departure times for particular buses (bus-level predictions).

​

• Every time new information about a bus location or weather is received in our system, these bus-level predictions are recalculated using the latest relevant model-level predictions.

Bus Level Predictions

Run Time �ML Models

Dwell Time �ML Models

A

B

C

D

E

F

Infrastructure Considerations

Speed

• Departure time predictions must be generated and served to customers in a timely manner as predictions become stale in minutes.
• The implementation method must be able to generate predictions every 90 seconds for up to a total 1200 buses (peak time) and their upcoming stops.

​

Storage

• Reporting database was growing at a rate of roughly 6TB a year.

​

Scaling

• With each bus line having an average of 30 segments and 30 stops, the system must be able to generate over 20,000 predictions per minute. This varies based on peak/non-peak times.

Infrastructure Implementation

Microsoft Azure cloud platform advantages:

• Easy to deploy resources, e.g. VMs, DBs
• Supports containers to help with scaling and deployment
• Technical support and consultation readily available
• Leverage Azure cybersecurity
• Options for disaster recovery/redundancy
• Matched the team skillset

​

Pilot Routes

13 pilot routes chosen based on the following factors:

• Trip length
• Number of stops
• Frequency
• Week day services
• Road types
• Vehicle types
• Bus rapid transit
• Weather sensitivity
• Bus type

​

​

Results

Machine Learning

Scheduled

RTTI

Blended

Actual

Time Of Day

Results

Results

Average Error

​

​

​

​

​

+/- 3 min 13 sec

+/- 1 min 41 sec

47.8% less error!

Average Unexpected Wait Time Per Boarding

​

​

​

​

​

60% less waiting per boarding!

+/- 3 min 27 sec

+/- 1 min 23 sec

Operational Support & Maintenance

Cloud cost reduction

• Optimize code to reduce VM usage
• Kubernetes to auto-scale VM cluster based on need
• Cheap cold-storage options for data archiving
• Reserved instances
• Turn-off/de-allocate unused resources

​

​

​

​

​

Operational Support & Maintenance

Drift Detection

• Like any ML model, predictions from our models will drift over time. The system must have the ability to automatically detect when models drift beyond an acceptable threshold and trigger model retraining with more recent data.

​

• With over 18,000 sets of machine learning models, models should have version control; model linage information must be traceable to allow future investigations and reproduction of models if required.

​

​

​

​

Operational Support & Maintenance

Drift Detection Workflow

​

​

​

​

Architecture

Highlights:

• The aggregator takes the weather and bus vehicle monitoring (VM) info to generate bus-level predictions
• Predictions queue “smooths” out the # of predictions
• Bus Schedule DB serves predictions to the customer
• Drift monitoring system ensures models are kept accurate

​

Project Timeline

Q2

2018

Q4

2018

Q1

2019

Microsoft Proof-Of-Concept on a single bus route

Design a more powerful ML algorithm

Implement ML predictions for 13 pilot routes

Q3

2019

Implement ML predictions for all remaining bus routes

Q2

2019

Include additional features to ML models

Q3

2018

Q4

2019

Implement drift detection and automated model creation & cost optimization

What’s Next?

• Improve ML algorithm with additional factors:
• Traffic
• Bus driver
• Special events

​

• Continue cost optimizations

​

• Monitor & improve data quality

​

Is Machine Learning Right For You?

• Do you have a lot of data?
• Is the data cleaned and validated?
• Do you have a business objective that is measurable with data?
• Does your team have the correct skills (programming & stats)?
• Is the necessary computing power & storage available to you?
• Do you have the ability to do drift detection?

​

If you answered “no” to any of these questions then you’re not ready for ML.

​

If you answered “yes” to all of these questions, then ML might be right for you!

Thank You!

https://www.translink.ca