Urban mobility generates massive amounts of data: vehicle locations, wait times, usage patterns, commuting preferences, payments. Most cities ignore them or collect them without utilizing them. Novoville Mobility aims to bridge this gap through Machine Learning and Artificial Intelligence at three levels.
WP3 of the project was dedicated exclusively to the analysis of AI/ML technological trends in mobility and laboratory simulations — laying the groundwork for the implementation of WP6.
The Three Levels of AI/ML in the Platform
Level 1 — Personalized Rewards (D6.5)
Algorithms that monitor each user's transaction history and automatically calibrate rewards based on the plan set by the administrator. A user who frequently uses a bicycle receives a different proposal than one who mainly uses parking. The logic is based on a Rule Engine with AND logic, streak multipliers (consecutive days of sustainable commuting), and flat bonuses for achievements.
Level 2 — Statistical Analysis of Commuting Behavior (D6.6)
Electronic reports that aggregate anonymized data to produce mobility KPIs per municipality. The system identifies trends, seasonality, geographical distribution of demand, and reward effectiveness. It feeds the administrative platform (D6.4) with dashboards for policy decisions.
Level 3 — Multimodal Routing
An optimal route algorithm that takes into account multiple means of transport simultaneously: walking, cycling, bus, parking. Optimization is based on three parameters: time, cost, environmental footprint. The user selects their priority and receives personalized recommendations.
The CO₂ Methodology — A Practical Example
One of the most interesting technical problems we faced concerns the calculation of CO₂ avoidance. How do you measure the amount of emissions that did not happen?
For public transport (buses)
Check-in only happens upon boarding — there is no check-out. Therefore, we do not know the route the passenger took. The solution: a multiplier coefficient per ticket, which encodes the average route distance and average duration for each line. It is calibrated per city and per ticket type.
For shared bicycles
Usage time is measured precisely (unlock–lock), but not the distance — GPS tracking is avoided for privacy protection reasons. The solution: a multiplier coefficient per minute of use, which estimates the average distance based on average bicycle speed. The comparison is made with the equivalent car journey.
Laboratory Simulations (WP3)
Before WP6 took over the development of production algorithms, WP3 verified the architectural approaches in a controlled laboratory environment. The simulations focused on:
- Clustering algorithms for grouping user mobility profiles
- Collaborative filtering for reward recommendations based on similar users
- Time-series analysis for demand forecasting in parking spaces and bike stations
- Graph algorithms for optimal multimodal routing
The results of WP3 confirmed the chosen architecture and directly fed into the specifications of WP6.
Privacy by Design
The entire AI/ML architecture has been designed with the principle of privacy by design: algorithms operate on anonymized or aggregated data. The exact GPS route is never stored. Behavioral data is processed locally wherever possible. Privacy protection is not an afterthought compliance — it is an architectural requirement.