Melbourne tram load estimation and real-time load prediction
This project aims to develop data fusion and machine learning models using passively and automatically collected data to estimate service utilisation for Melbourne trams, and to make real-time prediction of tram loads.
The outcome of the project will be adopted 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.
Project background
Compared to the conventional sources of travel demand data (travel surveys and strategic models), automatic and passively collected data (such as passenger counts, pedestrian sensors, and fare card transactions) has the advantage of coverage in three important dimensions: space, time, and population. This provides an unprecedented opportunity to monitor day-to-day variability of travel demand and understand demand responses to service disruptions, special events, restrictions (such as COVID-19) and operational interventions.
Now that the COVID-19 pandemic has changed the day-to-day lives of Australians and their travel decisions, service utilisation measures and information systems in this space has become more relevant. Public transport service utilisation changed during the pandemic, and travel demand patterns are expected to differ from the pre-pandemic distributions.
In the context of rapidly evolving demand patterns, automatic and passively collected data provides an opportunity to monitor demand patterns and service utilisation closely, continuously, and comprehensively; and it can provide decision support for efficient operational interventions.
Real-time crowding information can help passengers make path decisions COVID-safe decisions and improve their rides, as well as reduce crowding in the system. Real-time crowding data can also inform operational interventions such as procurement of supplementary and replacement service to alleviate excessive crowding, address service disruptions and improve service efficiency.
Despite the above-mentioned advantages, automatic data usually lacks the accuracy and depth of information required for planning and operational interventions. Fare card data typically include information on passenger origins, and in some cases destinations, as well as their time of travel and mode of ridership. Moreover, these datasets usually include fare card identifiers for trips taken within the same day and across days of observations.
However, attributes such as the actual origin, final destination, transfer locations, and travel purpose are usually missing in fare card datasets. In addition, due to the massive scale, passive nature, and automatic process of data collection, the possibility of errors and void data are higher than traditional survey methods.
As a result, modellers usually require special methods to impute missing or erroneous data. A growing body of literature in transport science and travel behaviour research is devoted to innovative and effective methods address fare card shortcomings for planning and operational applications.
Particularly relevant to public transport services in Melbourne, the free tram zone in the central business district is a blind spot (to the fare collection system) because passengers who travel within the free tram zone are not required to tap in (or out) their myki cards.
Despite the critically high passenger loads carried on the free tram zone services, vehicle load information is challenging to estimate due to the lack of direct observations. In addition, fare evading rides are also unobserved in myki data.
A large body of literature in public transit data science indicates that fare card data shortcomings can be effectively addressed for use in planning and operational applications, if complemented effectively by alternative data sources.
Trip chaining methods assume that users alight at the stop nearest to their subsequent boarding stop; to infer the alighting location for the last trip of the day, boarding location on next day is used. Recent research has indicated that data fusion techniques using alternative sources of data such as automated passenger count systems, weight and wi-fi sensors, and CCTV footage can complement fare card data to achieve reasonably accurate load profiles and crowding information.
Overall, the state-of-the-art methodologies in developing passenger flow models have proven effective and reliable for estimation of travel O-D matrices and passenger flows in multimodal PT networks, if data is complemented by the right alternative sources and the methodology is customised for the specific data shortcomings and specific project objectives.
Melbourne trams have significant connections with bus and train networks and multimodal trip chaining can be utilised to dig deeper and extract more information within the free tram zone.
In this project, we aim to use alternative and independent sources of data to complement fare card transactions for Melbourne tram service load estimation and crowding predictions. We will apply advanced trip chaining, network modelling and Machine Learning techniques (including deep learning and reinforcement learning) with alternative data sources (such as automatic passengers counts from available tram routes, on-board footage, train station turnstile counts, travel demand models and land use data, etc.) to infer reliable estimates and accurate short-term predictions for tram load profiles.
Project objectives
- Develop a passenger flow model
We will develop a model for passenger trajectory estimation, aggregate O-D matrix inference and service load estimation for tram services in Melbourne using the trip chaining method.
- Develop a calibration model to adjust for unobserved riders
We will develop a calibration model to correct the inferred O-D matrices and passenger load profiles. The calibration will apply to service load estimations to correct for fare evasion instances, unobserved trips, and trips within the free tram zone.
- Develop Machine Learning (ML) model for real-time prediction of service loads
We will develop data-fusion and machine models that builds on the historical load profiles to produce real-time predictions of service utilisation.
Please note …
Ongoing, this page will be a living record of this project. As it continues, matures, hits milestones, etc., we’ll add information, links, images, interviews and more. Watch this space!
