ACUMEN proposes a generic, privacy-preserving, data-driven modular digital paradigm for advanced network management, which aims at enabling efficient and reliable door-to-door journeys for people and goods, increased safety and resilience at the network level, and to make a critical contribution to achieving the transport goals set forth in the green deal. The main concept developed in ACUMEN is a modular, multi-layered Digital Twin (DT), a high-fidelity representation of integrated and interacting real complex systems, ultimately forming a digitised version of seamless and sustainable, connected urban mobility. This is complemented by plug-in modules, or digital tools, which represent the outcomes of the models (physics-based or data-driven based), data (including that generated via AI/ML approaches using said models), and simulation tools at the disposal of a city/road authority/mobility service provider. AI-powered digital tools supporting mobility management and decision-making, exploiting the modular DT architecture, will be developed by leading academic and research partners, in close cooperation with global industry partners and stakeholders. The DT platform will be demonstrated and validated through a set of comprehensive and carefully selected use cases, co-created with stakeholders, involving different scales and urban forms, to challenge the capabilities of ACUMEN with a diverse range of transport management problems and applications.
How We Contribute
As part of the ACUMEN project, the DSAIT team is involved in many tasks. Specifically, DSAIT has developed AI-based methods to efficiently process and analyze large volumes of data, enabling accurate and dynamic estimation of traffic states. Our research focuses on the automated identification of unexpected mobility patterns, integrating unsupervised learning techniques and trustworthy traffic conditions forecasting using theory-aware machine learning and hybrid approaches.
Regarding the road traffic incident identification, a bi-level framework is developed in order to (i) derive a daily (expected) traffic pattern and (ii) detect significant deviations from the expected pattern in real time. From the above process, a novel classification of traffic conditions also emerges, concerning the recurrency of the traffic conditions. In addition, theory-driven , short-term deep learning traffic forecasting models are developed, taking account of the distance of the predictions from the fundamental diagram, offering enhanced performance and reliability, which together with traffic state identification, converge to support the development of a comprehensive risk assessment framework. This framework actively involves human decision-making, automatically detecting situations where human intervention is required.
To this end, we have developed a human-in-the-loop system, equipped with an intuitive user interface designed for ease of use by all operators. This system ensures smooth integration into the overarching vision of the project, aiming for seamless multimodal network and traffic management.
