Highlights Learning-based method for traffic signal restoration sequence ordering. Modified Structural Recurrent Neural Network (SRNN) for predicting transportation system performance. Compared Step-LR and Cosine-Annealing-LR strategies in terms of the fine-tune of learning rate. Demonstrated the method with numerical experiments.

    Abstract Traffic signal failures could result in significant local and system-level performance degradation. Sequencing the restoration of failed signals with limited resources is a challenging problem—capturing dynamic changing transportation system performance following a feasible solution requires tedious computation, and the short time frame for restoring failed signals makes these decisions time-sensitive and should be determined in a timely manner. Feasible Signal Restoration Sequence Ordering (SRSO) problem, as a critical building block to optimize the restoration sequence, has not been well-studied in the existing literature, nor have solutions to address the computational burden issue. In this work, a machine learning model based on Structural Recurrent Neural Network (SRNN) is proposed to predict system performance, i.e., aggregated accumulated total delay, following a given restoration sequence, to address the computational burden of the simulation-based performance evaluation. Spatio-temporal (ST) graph representation is leveraged in this methodology to take the topological information, i.e., how adjacent movements interact with each other, into consideration. Although microscopic simulation is used to obtain the ground truth performance, which is still time-consuming for a group of signal failure scenarios, a trained machine learning model can surrogate the tedious computation in the decision-making process in a timely manner. The challenges to build this machine learning model effectively and efficiently are two-folds. First, a transportation system is a typical dynamic system whose behavior is constrained by the topology of the network. Therefore, both spatial and temporal interactions between road sections should be captured to predict system performance effectively. Second, in the context of signal failure, system performance is exposed to a disrupted control strategy, which makes it even more challenging to predict system performance effectively. The original SRNN model could address the first challenge, and the movement feature representation and its integration to the SRNN model proposed in this work could address the second. A case study was conducted to demonstrate the operability and effectiveness of the proposed methodology. It is demonstrated that the signal restoration sequence could impact system performance during and after the restoration process significantly. Then, both Aggregated Accumulated Total Delay (AATD) prediction accuracy and the performance of restoration sequence ordering is evaluated for the case study network. Outcomes of the case study show that the learning-based model can help identify sequences ranked in top 8% of optional sequences referring to ground truth information. Furthermore, in terms of the fine-tune of the learning rate, the Cosine-Annealing-LR strategy leads to both lower loss value, better ordering performance, and shorter delay experienced in the restoration process, compared to the Step-LR strategy.


    Zugriff

    Zugriff prüfen

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    Learning-based restoration sequence ordering for multi-site traffic signal failure


    Beteiligte:
    Zhao, Tingting (Autor:in) / Zhang, Yu (Autor:in)


    Erscheinungsdatum :

    2021-12-08




    Medientyp :

    Aufsatz (Zeitschrift)


    Format :

    Elektronische Ressource


    Sprache :

    Englisch




    Traffic Signal Sequence Alert

    PRIEST III CHARLES WILLIAM | Europäisches Patentamt | 2016

    Freier Zugriff

    Intelligent traffic signal failure detection circuit

    HU HONGSHENG | Europäisches Patentamt | 2015

    Freier Zugriff

    Multi-intersection traffic signal control method based on deep reinforcement learning

    DENG HENG / WANG YULONG / GAO YANG et al. | Europäisches Patentamt | 2024

    Freier Zugriff


    Reinforcement learning-based multi-agent system for network traffic signal control

    Arel, I. / Liu, C. / Urbanik, T. et al. | Tema Archiv | 2010