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

Takao Dantsuji,Dong Ngoduy,Ziyuan Pu,Seunghyeon Lee,Hai L. Vu

2024-08-13

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.

Systems and Control

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

This paper aims to address the challenges of real-time traffic demand calibration (OD demand matrix calibration) in the face of sudden events (such as accidents, extreme weather, etc.). Existing urban traffic management systems are often slow to respond to these sudden events and cannot quickly adapt to rapidly changing traffic conditions. Specifically, the paper proposes a hybrid neural network architecture for real-time calibration of OD demand matrices to improve the accuracy and reliability of traffic simulations, thereby better coping with both recurring and non-recurring traffic conditions.

### Background and Motivation

1. **Limitations of Existing Systems**:
- Existing automated urban traffic management systems are primarily designed to alleviate traffic congestion and reduce emissions, but they perform poorly in responding to sudden events.
- These systems are usually passive, only reacting after an event occurs, leading to delayed response times and inefficiency in handling sudden changes.

2. **Importance of Real-Time Traffic Simulation**:
- Real-time traffic simulation is an important tool for dealing with non-recurring traffic conditions, especially microscopic traffic simulators that can describe the behavior of individual vehicles in detail.
- By utilizing real-time data and continuous parameter updates, these simulations can provide accurate short-term traffic condition forecasts, aiding in the formulation of proactive traffic management strategies.

3. **Challenges of OD Demand Matrix Calibration**:
- The OD demand matrix is a key input parameter in traffic simulation, with each cell representing the number of trips from one specific area or point to another.
- Due to the inherent variability, randomness, and unpredictability of traffic demand, accurately calibrating these matrices under non-recurring traffic conditions is very challenging.
- Traditional calibration methods rely on historical traffic data, which may fail to capture sudden changes in driver behavior or route choices caused by sudden events.

### Solution

The paper proposes a hybrid neural network (NN) architecture for real-time OD demand calibration. The main features of this framework include:

1. **Data Integration**:
- Seamless integration of traffic data from limited real-time data and rich simulation environments to improve the accuracy and reliability of traffic predictions.

2. **Real-Time Parameter Updates**:
- Using a meta-model-based backpropagation method to achieve real-time updates of NN parameters, minimizing the discrepancy between actual and simulated traffic conditions.

3. **Offline Pre-Training**:
- Improving computational efficiency through offline pre-training with a meta-model.

4. **Iterative Optimization**:
- Dynamically adjusting traffic patterns through continuous real-time iterations to adapt to unpredictable traffic disruptions.

### Validation and Application

The paper validates the model's performance under different disruption scenarios through a toy network and a real case study of the Tokyo highway corridor. The results demonstrate that advanced machine learning techniques have great potential in developing proactive traffic management strategies, significantly outperforming traditional passive systems.

### Main Contributions

1. **Combining Model-Driven and Data-Driven Approaches**:
- Model-driven approaches utilize domain knowledge (traffic models) to provide structured and interpretable predictions but have limited scalability and real-time parameter calibration capabilities.
- Data-driven approaches, while scalable and capable of capturing complex patterns, require extensive retraining under rapidly changing traffic conditions, making real-time adjustments difficult.
- The hybrid ML framework proposed in the paper combines the advantages of both, using traffic simulation models as the digital twin of the network to enhance the robustness of traffic flow behavior.

2. **Real-Time Adaptability**:
- Generating additional synthetic data through real-time simulation, continuously adjusting the learned algorithms to quickly adapt to environmental changes.

3. **Computational Efficiency**:
- Reducing computational complexity through offline pre-training with a meta-model, providing a good starting point for real-time simulation.

In summary, the method proposed in this paper performs excellently in dealing with non-recurring traffic conditions, offering new solutions for urban traffic control and management.