Estimating and predicting transit ridership
By Moov AI | Published on April 8, 2022
Problem and objectives
Problem
- Starting in March 2020, the development of telecommuting solutions by several Quebec companies rapidly and significantly reduced the number of passengers using the Montreal metro. Some users, particularly those with weakened immune systems, had concerns about their safety within the metropolitan network. The implementation of preventive measures to encourage physical distancing in public transport proved necessary.
- To address these issues, the STM wanted to develop a tool to estimate ridership in real time and predict the number of passengers in the metro cars. This would allow the STM to facilitate the physical distancing inside the cars, increasing the feeling of security for the passengers.
Objectives
- Estimate the number of passengers in real time
- Predict the number of future passengers
Proposed solution
Solution
From June to December 2020, Moov AI led the design and development of a project to estimate ridership in the subways of the Société de transport de Montréal (STM). The project was carried out in an accelerated mode due to the outbreak of the COVID-19 pandemic.
Principle
Using the scales built into the Azur metro cars, Moov AI first developed an algorithm to estimate the number of people present in the wagon in real time, within 15-minute intervals. Then, the collected estimates were used in the building of a deep learning model, allowing to predict the ridership of the following days. Finally, Moov AI created a programming interface and the infrastructure required to run the deep learning algorithms.
Role of AI
Due to the complexity and multi-dimensional nature of the subway network, which includes multiple platforms and cars, each with a unique configuration, an AI-centric solution was required. Without AI and an algorithm that could react to the multiple signals sent to it, it would have been impossible for Moov AI to achieve a high degree of accuracy.
Techniques used :
- Algorithms;
- Machine learning;
- Deep learning (multilayer perceptron) ;
- Deep neural networks;
- Encoders;
- Decoders;
- Regression ;
- Active online learning model;
- Ensemblistic learning (XGBoost);
- Recurrent deep learning (LSTM)
Impacts
Main outcomes
Impacts on the STM
The project has had a significant impact on modernizing the technology platform, its infrastructure, its deep learning algorithms and the advancement of AI knowledge at the STM. They now have the tools to better adapt to changing health situations and can use the data received to support their decision making. The results of the models are used to disseminate information, so that passengers feel safer to travel by metro at all times.
During this project, the STM has deepened its expertise in the maintenance of AI-enabled infrastructure. Based on the training it received, the Montreal transportation company has perfected its knowledge of the deep learning models developed by Moov AI to the point of reaching a complete level of autonomy.
Impacts on Moov AI
At Moov AI, the mandate was a catalyst for further business development. Having delivered the solution in six months, the results of this project provided them with significant exposure to organizations looking for similar solutions. Internally, the project allowed Moov AI to enrich their mixed team experience (Moov AI-STM) and improve their processes for deploying deep learning projects.
Generation and dissemination of new knowledge
Moov AI contributed to technology transfer by training two members of the STM team. The knowledge acquired has allowed them to maintain and improve the AI models, with which they now have complete autonomy.
Challenges takled
As part of the challenges met during this project, the scales were sensitive to the speed of the cars and the curves they were traveling on. Due to this sensitivity, the data collected by the scales varied constantly. These variables were then considered in the deep learning algorithms.
In addition, it was important for Moov AI to consider standards and regulations when developing a project with a public transportation agency. The company had to comply with STM regulations and use standardized tools.
The accuracy of the results was an important factor, as the impact of a false negative can be considerable, i.e., an estimate that suggests that there are only a few people inside the metro car, when in fact the passenger limit is exceeded. For instance, a situation we wanted to avoid was for a person with a weakened immune system to take the metro at a time of heavier traffic, after consulting the STM information. Therefore, Moov AI put more effort into designing a solution that exceeded the accuracy target set in the design phase of the project. The deep learning model that Moov AI developed was able to estimate the number of users per car to within three people.
Conditions for success
Team mobilised
Internal Resources
- Data analysts;
- Solution architects;
- Data scientists;
- Data Engineers;
- Software developers;
- UI/UX designers;
- AI project manager.
External Resources
- STM-Moov AI hybrid team
- Project managers
Project stages
Organized a one-day workshop to understand the problem and establish the strategy;
Designed a minimum viable model (model design without infrastructure connections);
Developed the machine learning infrastructure;
Designed the minimum viable product (model connection to infrastructure);
Workflow improvements;
Delivery of documentation and code to STM.