Melbourne trams: Better data to improve operation and planning

The Melbourne tram load estimation and real time load prediction project, undertaken by the Department of Transport and Planning Victoria, Cubic Transportation Systems, and the University of Melbourne has been completed and is available for download below.

This project’s aim was to improve the efficiency of tram operations, to assist with route and service planning, integration, and operation of tram routes with other modes of transport.

Objectives

In strategic transport planning, a basic understanding of regular travel demand volumes in zone aggregated levels are usually sufficient, which can be accessed by the four-step travel demand forecasting models, or similar models.

However, for operational applications, demand variability information is becoming increasingly relevant and essential. The transit industry has recently started to use demand information to support data-driven decision making. For example, transit operators can adjust service frequency to accommodate for variability in passenger demand, develop demand-responsive service to better target areas with low passenger demand, inform customers of the crowding levels and assist them in their journey planning to avoid service overcrowding.

The development of these strategies requires an understanding of more detailed passenger demand information with day-to-day and within-day variability compared to strategic level models. It is particularly important to understand demand responses to service disruptions, special events, restrictions (such as COVID-19) and operational interventions (“before/after” analysis).

Providing transit operators with demand patterns during these events could enhance their ability to respond to uncertainty and provide customers with more reliable and efficient services and information. The conventional types of demand data from on-board surveys are labour-intensive and costly. As a result, the sample sizes are inevitably small for operational applications.

Myki data is structured around the current fare policy and this presents four major challenges when using the data to make inferences around patronage:

1. Cleaning the myki dataset as relevant for patronage estimation
2. Inferring stop number for a transaction where myki returns an area encompassing multiple stops
3. Imputation where no transaction was recorded for a boarding/alighting
4. Chaining transactions together into a trip-chain

The number of boardings/alightings that need to be imputed was high. This was especially true for trips starting in the FTZ where tapping on or off is not required resulting in gaps in the data.

This project consisted of three tasks. Task 1 utilised Myki data to estimate the passenger flow volumes between all identified origin and destination pairs.

Task 2 first analyses the distribution of missing Myki data and uses an econometric model to demonstrate how, when, and where the missing data occurs. Then a suite of passenger load estimation models are developed, based upon the Myki origin-destination matrix (OD) a collection of diverse datasets, including Automated Passenger Count (APC) data, and automatic vehicle location data (AVL), and other public data sourcesTask 3 focused on building a prediction model to utilise data-trained understanding of the public transport network utilisation to predict the expected service loads in real-time.

Tram route 96 provided a special challenge because this route travels through the FTZ and outbound to the south. Myki data for this route in particular was subject to both missing observations and underestimation

Task 1: Passenger flow model

The most commonly collected data includes passenger origin, day, date and time of travel, with these basic datasets providing fare identifiers that can form the basis for evaluation across same day, weeks, or years of observations. For this task, data from the Myki fare collection system was examined and evaluated.

Researchers broke down Myki trips into individual journeys and associated them with unique origin-destination (OD) pairs.

Four major challenges were identified as part of this work:

1. Correction of dataset issues and improbable entries;
2. Inference of location at which Myki transactions occur;
3. Imputation of missing or correction of erroneous Myki transactions; and
4. Correctly chaining transactions for individuals inferred destinations and transfer locations.

Conclusions

This project investigated and illustrated how existing data sources for service load estimations could be augmented with machine learning models and alternative data sources.

The touch-on ratio for Melbourne trams is subject to significant variations over space and time, and by using data from the Myki system alone results in missing data, resulting in inaccurate service and tram load predictions. Using a simple expansion factor does not work for estimations and predictions here because of the high variation in passenger loads.

Data gaps are reduced when APC systems data are employed in conjunction with Myki data and econometric analysis. Predictions are highly accurate when employing service performance variables such as dwell time and headway, however without these variables prediction accuracy is challenging because of the highly dynamic nature of tram loading.

Unfortunately, not all Melbourne transit routes are equipped with APC systems.     For routes without APC devices, the researchers posit that it will be necessary to use machine learning models trained from other routes. This is possible if 10% vehicles of every service route were to be equipped with APC systems.

Recommendations

For higher accuracy and prediction rates, the researchers also assert it is essential to employ a feedback loop during prediction periods with accuracy improved again if there were more overlapping stops between training and test routes.

There is further work to be done here as no information has been gathered from tram lines in the north-south direction. These routes pass through the CBD’s TFZ and continue outbound. Researchers flag priority for APC sensor installation should be given to these tram routes for future evaluation.

The work done here illustrates that expansion of operational data and increased accuracy in predictions will ultimately help both transit operators and passengers.

Expected project impacts

The practice of transport engineering has always been challenged by the lack or shortage of proper data, and this gap has always been compensated by intense modelling and simulation, always prone to strong assumptions and lack of precision.

This situation has recently changed. Individual’s mobility data is now collected passively from different sources. Our research, aims to utilise data science methodologies and explore possibilities to extend the traditional modelling and simulation approaches to improve the efficiency and sustainability of urban transport systems.

See also an academic paper published from this project: Transferable supervised learning model for public transport service load estimation.

Dr Neema Nassir, Senior Lecturer in Transport Engineering, University of Melbourne

Final report

Download your copy of the final report, Service Load Estimation and Real-Time Crowding Prediction for Melbourne Trams, by clicking the button below.