Download to read offline
![Scaling-up Approach - Prediction of Uptake
1. Prediction of average daily km recorded
by mode for each participant for each
district in Austria
○ Same incentive regime as AI-CENTIVE pilots assumed
Scenario Number of persons participating
companies communities
S1 54 718 27 623
S2 136 795 95 122
S3 547 182 239 039
Average daily distance [km] recorded by participants
walking cycling public transport ridesharing passenger
company member 0,22 0,78 0,67 1,45
community member 0,34 0,54 0,27 0,75
2. Prediction of participants
○ Companies with CSRD reporting requirements main
targets (> 250 staff) for rollout
○ 3 rollout scenarios: 10%, 25%, 100% participate
○ 30% of employees of those participate
○ residents in municipalities proportionally](https://image.slidesharecdn.com/master-ppai-centivefinalstakeholdermeeting20251119-251125083945-7a335208/75/AI-CENTIVE-Final-Results-Final-Stakeholder-Meeting-35-2048.jpg)
AI-CENTIVE Final Results - Final Stakeholder Meeting
- 1. Final AI-CENTIVE StakeholderMeeting Nov 19 2025 | 4-6 pm | Vienna Presentation of the project results and achievements
- 2.
- 3. Time What Who 16:00-16:05Welcome Emilia Schützenhofer, DIO 16.05-16.20 AI-CENTIVE: project highlights Lyndon Nixon, MODUL Technology | AI-CENTIVE project leader 16.20-16.35 An app for sustainable mobility incentives Dave Kock, ummadum 16.35-16.50 AI improved weather prediction Toni Jurlina, Geosphere 16.50-17.05 A dashboard for sustainable mobility insights Arno Scharl, webLyzard 17.05-17.20 Impacts on mobility behaviour and sustainability Astrid Gühnemann, BOKU 17.20-17.30 Short break 17.30-18.00 Panel: wrap up on how we can promote more sustainable mobility in Austria in the future All project partners 18.00 Networking Agenda
- 4. Project Highlights Lyndon Nixon MODULTechnology Project leader
- 5. Mission & Vision Sustainablemobility behaviour is a difficult goal to reach, as people are not willing to easily change their habits just because it may help the environment. • Our mission is to develop AI- based incentivisation techniques for influencing citizen’s mobility choices based on multimodal models of mobility activity and data analytics. • Our ultimate vision is to enable and incentivize Austrian citizens to find more sustainable mobility choices, increasing awareness and affecting public opinion to develop a more positive attitude towards those sustainable mobility choices.
- 6. 1. Citizens makemany mobility trips every week, often using less sustainable options (e.g. private car) even though more sustainable options may be present (e.g. carsharing, public transportation, bicycle or walking) 2. CO2 emission reduction is broadly known to be a desirable goal but appears insufficient to make people actively choose more sustainable options when making a trip 3. Would rewards for sustainable behaviour incentivise the selection of more sustainable mobility forms when making a trip? Which ones? What extent of impact do they have on the “default” mobility behaviour? Incentives and sustainable mobility
- 7. - create acommunity in the Ummadum app - assess default mobility behaviour via online survey - use AI and weather to predict future trip plans based on past trip patterns - offer different incentives / rewards before the next trip - analyse changes in trip behaviour as logged in the app by the users - assess environmental impact from more sustainable mobility behaviour - develop strategies for broader incentivization in Austria to generate even more sustainable mobility decisions AI-CENTIVE: what we did
- 8. Best performing approacha combination of GCN and machine learning Patterns used past 6-12 weeks of user trip data Choice of notifications for the user made based on model’s confidence score (in its own prediction) For the stakeholder, we added explainability of model inferences Hybrid incentivization approach
- 9. Classic notifications: • Incentivizeparticipant based on previous sustainable mobility • Goal: Reward and motivate participants for their activities AI recommendations: • Recommend to participant future sustainable mobility • Use contextual information (behaviour, mobility preferences, location, time, ...) • Goal: (Re-)activation of participants AI weather notifications: • Recommend future mobility including weather information • Goal: Provide added value by giving relevant weather data Notifications
- 10. ● two pilotsin 2024 and 2025 to incentivize ummadum users ○ mobility behaviour and context modelling ○ empirical study on the impact of incentives ● AI improved weather prediction ● A Web based dashboard for stakeholder insights (mobility and incentives) ● Assessment of the sustainability impact (of sustainable mobility incentivization) Pilots, results and lessons learnt Time What Who 16.20-16.35 An app for sustainable mobility incentives ummadum 16.35-16.50 AI improved weather prediction Geosphere 16.50-17.05 A dashboard for sustainable mobility insights webLyzard 17.05-17.20 Impacts on mobility behaviour and sustainability BOKU
- 11. An app forsustainable mobility incentives Dave Kock
- 12.
- 13.
- 14. ummadum App Activities - Usingmultiplier: Events, Companies and Cities - Creating impact & usage thru Recommendations - Monthly Challenges
- 15.
- 16. Realtime forecast andobservational data provision IFS-DET forecast data (global model from ECMWF) - Location-specific 10-day forecasts through bi-linear interpolation from model grid - Hourly data for all districts (zip codes), once per week (Mondays) - Parameters: 2-m temperature, relative humidity, 10-m wind speed and direction, total cloudiness, precipitation rate, global radiation and a weather symbol INCA gridded observations - Operational analysis and nowcasting system at Geosphere Austria, operating at 1-km horizontal resolution - Combines radar, satellite, station and NWP data - Parameters: 2-m temperature, relative humidity, 10-m wind speed and direction, total cloudiness, precipitation rate, global radiation and a weather symbol - Hourly data interpolated onto all districts (zip codes), once per week for the last week
- 17. Advancements in Statisticaland Machine Learning Ensemble Postprocessing Primary context Ensemble forecasts are primary source of extended-range weather information Role of Statistical Calibration Improves reliability of ensemble forecasts Trend in Statistical Models Shift toward flexible distributions to better represent forecast uncertainties Challenges with Flexible Models Parameter estimation and modelling remain complex and computationally intensive Emergence of Machine Learning Methods Regression trees and random forests increasingly used for ensemble calibration Inspiration from Neural Network Neural networks can efficiently estimate distribution parameters Enable highly adaptable-probabilistic postprocessing Proposed Method - Quantile Function Regression (QFR) Parameter estimation and modelling remain complex and computationally intensive
- 18. Quantile Function Regressionand Bernstein Polynomials Limitations of Classical Quantile Regression Dependency between observations and ensemble is modeled only for a finite set of quantiles Why Bernstein Polynomials Matter in QFR Used as basis functions for quantile modeling due to their smoothness, non-negativity and boundedness From Traditional Postprocessing to Neural Networks Earlier methods: MOS-based corrections, tree-based regression ML models Modern deep learning enables flexible distribution learning Bernstein Quantile Networks (BQN) A neural approach that models predictive quantile distributions via Bernstein polynomials What it does overall Defines and trains a BQN for flexible quantile regression Maps input features to a quantile function Captures entire conditional distribution to the response variable Programming language Julia / Flux package Originally developed at MET Norway Adapted and tested at GeoSphere Austria Data Processing • A single BQN model was trained, using: Selected NWP variables interpolated at each point Mean precipitation (NWP) at a point level Leat time treated as a continuous variable
- 19. Training Workflow andTime Periods Procedure Overview - Preparation of training data - Retrieval of forecasted and observed point data - Processing into 12h accumulations - Training of BQN models - Inference performed at station/grid-point level Time Window - Dataset covers the year 2023 for training and validation - Training period: January - July 2023 - Validation period: August - September 2023 - Test period: October - December 2023 - Model later tested on September 2024 flood event as well
- 20. Flood event Boriscase study Key findings precipitation - Both global and regional NWP models performed remarkably well - The predictions are characterized by 1) unusual low variability from run to run, 2) unusual low variability from model to model and 3) unusual low variability wrt to the location of the absolute peak values Wind evaluation - Wind speeds during Boris high, but climatologically not exceptional - Similar to precip, the models performed considerably well - Highest skill for the 1km ensemble model in terms of SEDI
- 21. Flood event Boriscase study Demonstrated Improvements ML-based post-processing (BQN) models are able to improve the NWP forecasts at local scales across Austria Enhanced performance on a both time period and a case study Future Development Directions Train BQN on additional meteorological variables Extend to multiple NWP model, including ensemble inputs Compare against purely data-driven fc models
- 22. Visual Analytics Dashboardfor Sustainable Mobility Insights Arno Scharl
- 23.
- 29. Impacts on mobilitybehaviour and sustainability Astrid Gühnemann (Reinhard Hössinger, Tobias Dürrhammer, Valerie Jeepjua)
- 30. Research tasks andApproach Tasks and Research Questions 1) Empirical study on the impact of incentives: •For which incentives can a (significant) change in user behaviour be observed? •How strongly do users respond to different incentives? •Are there differences between modes? •Are there differences between user groups? •Do AI generated incentives have a stronger impact than rule-based incentives? 2) Sustainability impact assessment: •How big is the potential for change towards sustainable modes in Austria if the pilot results were scaled up? •If changes have been observed, how big is the impact on key sustainability criteria, particularly CO2 2 1
- 31. Data bases -Ummadum Data and User Survey • 2 survey waves (Jan-Sep 24, Oct 24 - Mar 25) • 928 valid answers (74% of all users, 79 % of active users) • provided socio- demographic data on users • full-time employees and urban citizens overrepresented
- 32. • Each ofthe five travel modes modelled separately • Observation units = user-days by travel mode • Non-users of respective mode were excluded (at least one activity with this mode required) • Dependent variable = binary response ○ no activity recorded (0) ○ one or more activities recorded (1) • Model type = binary logit estimated in R using package Apollo (Hess and Palma, 2019) • Modelling Approach
- 33. Model Results Impactsof Incentives - Comparative Effect ∑=27
- 34. Sustainability Assessment -Framework ● Identification of relevant impacts and indicators ● Developed in participatory logic mapping exercise Logic Map excerpt:
- 35. Scaling-up Approach -Prediction of Uptake 1. Prediction of average daily km recorded by mode for each participant for each district in Austria ○ Same incentive regime as AI-CENTIVE pilots assumed Scenario Number of persons participating companies communities S1 54 718 27 623 S2 136 795 95 122 S3 547 182 239 039 Average daily distance [km] recorded by participants walking cycling public transport ridesharing passenger company member 0,22 0,78 0,67 1,45 community member 0,34 0,54 0,27 0,75 2. Prediction of participants ○ Companies with CSRD reporting requirements main targets (> 250 staff) for rollout ○ 3 rollout scenarios: 10%, 25%, 100% participate ○ 30% of employees of those participate ○ residents in municipalities proportionally
- 36. • Modal shiftrates from a recent study from the UK (Arriagada et al., 2025) to calculate car-km saved: • -7.30% of walking trips replaced a car trip • -6.08% of cycling trips • -3.75% of PT trips • Ridesharing: based on commuting mode shares Predicted Sustainability Impacts • Calculation of CO2 savings (average emission rates) and increase in physical activity through AI-Centive rollout
- 37. • Carefully craftedincenctives can support the use of sustainable modes • Well-designed challenges with high chance for any reward are particularly effective • "Umweltverbund" respond more to gamification elements, ride-sharing more to economic benefits • AI can particularly support the delivery of targeted and timely reminders that keep users engaged • Long-term user engagement is essential for sustaining the app business • Continuous user engagement is a requisite for establishing and maintaining habits of sustainable mode choice • Sustainability impact of "soft" measures such as incentives are known to be limited on their own but important to accompany "hard" measures (e.g. better infrastructure, parking restrictions ...) Conclusions
- 38.
- 39. How can wepromote more sustainable mobility in Austria in the future?
- 40.