Dynamic Origin-Destination Matrix Estimation Using Metamodel-Based Model Predictive Control for Real Time Application

This paper proposes a method to advance real time calibration methods for microscopic traffic simulations. Conventional dynamic origin-destination matrix estimation (DODE) methods utilize metamodels to achieve computational advantages. However, they may estimate inappropriate origin-destination (O-D) matrices because of inherent differences between metamodels and simulations. An optimal control problem formulation is proposed for the DODE problem, and a metamodel-based model predictive control (MPC) approach that considers the uncertainty of the metamodel. The primary inspiration for the proposed method is to control the input O-D matrix while monitoring, at every time step, how the input demand influences the traffic system across multiple time steps. A data-driven metamodel based on an attention mechanism is developed to represent the dynamics of the microscopic traffic simulation. In addition, the metamodel is combined with one of the MPC algorithms, a model predictive path integral. The experiments were conducted on the Nguyen–Dupuis network and in Siheung City, South Korea. Compared with benchmark methods on each target network, errors were reduced by 35.3% on the toy network and 26.2% on the real-world network. This approach exhibits robust calibration performance against misestimated O-D matrices while considering the uncertainty of the metamodel. This paper presents a novel form of an O-D matrix estimation method for real time calibration that complements the existing stream of literature and provides a basis for developing real time simulations.