A hybrid neural network for real-time OD demand calibration under disruptions

TL;DR

This paper tackles real-time OD demand calibration under disruptions to improve the fidelity of microscopic traffic simulations. It introduces a hybrid neural network that couples real-time observations with a traffic model digital twin, enabled by a differentiable metamodel-based backpropagation and offline pre-training to update OD demands ${D}^{ij}_{t+1}$. The framework uses a tractable metamodel $h({\bf D}_{t+1})$ to propagate gradients, allowing online adaptation of NN parameters ${\theta}_t^*$ as conditions change. Case studies on a toy SUMO network and the Tokyo expressway corridor demonstrate improved OD predictions and robust adaptation under both recurrent and non-recurrent disruptions, highlighting potential for proactive traffic management with real-time data assimilation.

Abstract

Existing automated urban traffic management systems, designed to mitigate traffic congestion and reduce emissions in real time, face significant challenges in effectively adapting to rapidly evolving conditions. Predominantly reactive, these systems typically respond to incidents only after they have transpired. A promising solution lies in implementing real-time traffic simulation models capable of accurately modelling environmental changes. Central to these real-time traffic simulations are origin-destination (OD) demand matrices. However, the inherent variability, stochasticity, and unpredictability of traffic demand complicate the precise calibration of these matrices in the face of disruptions. This paper introduces a hybrid neural network (NN) architecture specifically designed for real-time OD demand calibration to enhance traffic simulations' accuracy and reliability under both recurrent and non-recurrent traffic conditions. The proposed hybrid NN predicts the OD demand to reconcile the discrepancies between actual and simulated traffic patterns. To facilitate real-time updating of the internal parameters of the NN, we develop a metamodel-based backpropagation method by integrating data from real-world traffic systems and simulated environments. This ensures precise predictions of the OD demand even in the case of abnormal or unpredictable traffic patterns. Furthermore, we incorporate offline pre-training of the NN using the metamodel to improve computational efficiency. Validation through a toy network and a Tokyo expressway corridor case study illustrates the model's ability to dynamically adjust to shifting traffic patterns across various disruption scenarios. Our findings underscore the potential of advanced machine learning techniques in developing proactive traffic management strategies, offering substantial improvements over traditional reactive systems.

Figures (18)

• Figure 1: A hybrid neural network for real-time simulation
• Figure 2: The proposed real-time simulation framework
• Figure 3: The integration of DNN with traffic simulator
• Figure 4: The architecture of metamodel-based backpropagation
• Figure 5: The toy network and its demand profile