An adaptive, fault-tolerant system for road network traffic prediction using machine learning

• Autores: Rafael Mena Yedra
• Directores de la Tesis: Ricard Gavaldà Mestre (dir. tes.) , Jordi Casas Vilaró (codir. tes.)
• Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
• Idioma: español
• Texto completo no disponible (Saber más ...)
• Resumen

  • This thesis has addressed the design and development of an integrated system for real-time traffic forecasting based on machine learning methods. Although traffic prediction has been the driving motivation for the thesis development, a great part of the proposed ideas and scientific contributions in this thesis are generic enough to be applied in any other problem where, ideally, their definition is that of the flow of information in a graph-like structure. Such application is of special interest in environments susceptible to changes in the underlying data generation process. Moreover, the modular architecture of the proposed solution facilitates the adoption of small changes to the components that allow it to be adapted to a broader range of problems. On the other hand, certain specific parts of this thesis are strongly tied to the traffic flow theory.

    The focus in this thesis is on a macroscopic perspective of the traffic flow where the individual road traffic flows are correlated to the underlying traffic demand. These short-term forecasts include the road network characterization in terms of the corresponding traffic measurements –traffic flow, density and/or speed–, the traffic state –whether a road is congested or not, and its severity–, and anomalous road conditions –incidents or other non-recurrent events–. The main traffic data used in this thesis is data coming from detectors installed along the road networks. Nevertheless, other kinds of traffic data sources could be equally suitable with the appropriate preprocessing.

    This thesis has been developed in the context of Aimsun Live –a simulation-based traffic solution for real-time traffic prediction developed by Aimsun–. The methods proposed here is planned to be linked to it in a mutually beneficial relationship where they cooperate and assist each other. An example is when an incident or non-recurrent event is detected with the proposed methods in this thesis, then the simulation-based forecasting module can simulate different strategies to measure their impact. Part of this thesis has been also developed in the context of the EU research project "SETA" (H2020-ICT-2015).

    The main motivation that has guided the development of this thesis is enhancing those weak points and limitations previously identified in Aimsun Live, and whose research found in literature has not been especially extensive. These include: • Autonomy, both in the preparation and real-time stages.

    • Adaptation, to gradual or abrupt changes in traffic demand or supply.

    • Informativeness, about anomalous road conditions.

    • Forecasting accuracy improved with respect to previous methodology at Aimsun and a typical forecasting baseline.

    • Robustness, to deal with faulty or missing data in real-time.

    • Interpretability, adopting modelling choices towards a more transparent reasoning and understanding of the underlying data-driven decisions.

    • Scalable, using a modular architecture with emphasis on a parallelizable exploitation of large amounts of data.

    The result of this thesis is an integrated system –Adarules– for real-time forecasting which is able to make the best of the available historical data, while at the same time it also leverages the theoretical unbounded size of data in a continuously streaming scenario. This is achieved through the online learning and change detection features along with the automatic finding and maintenance of patterns in the network graph. In addition to the Adarules system, another result is a probabilistic model that characterizes a set of interpretable latent variables related to the traffic state based on the traffic data provided by the sensors along with optional prior knowledge provided by the traffic expert following a Bayesian approach. On top of this traffic state model, it is built the probabilistic spatiotemporal model that learns the dynamics of the transition of traffic states in the network, and whose objectives include the automatic incident detection.