Traffic Modeling and Forecast based on Stochastic Cell-Automata and Distributed Fiber-Optic Sensing -- A Numerical Experiment

TL;DR

This paper presents a novel approach combining stochastic cell-automata and distributed fiber-optic sensing for accurate real-time traffic modeling and short-term forecasting, outperforming traditional point sensors.

Contribution

It introduces an optimal parameter estimation method using particle filters and DFOS data, enhancing traffic forecast accuracy and microscopic traffic state estimation.

Findings

Successful derivation of model parameters from DFOS data

Traffic forecast error of approximately 10 km/h for 60-minute prediction

DFOS outperforms point sensors in traffic modeling accuracy

Abstract

This paper demonstrates accurate traffic modeling and forecast using stochastic cell-automata (CA) and distributed fiber-optic sensing (DFOS). Traffic congestion is a dominant issue in highways. To reduce congestion, real-time traffic control by short-term forecast is necessary. For achieving this, data assimilation using a stochastic CA model and DFOS is promising. Data assimilation with a CA enables us to model real-time traffic flow with simple processes even when rare or sudden events occur, which is challenging for usual machine learning-based methods. DFOS overcomes issues of conventional point sensors that have dead zones of observation. By estimating optimal model parameters that reproduce observed traffic flow in the simulation, future traffic flow is forecasted from the simulation. We propose an optimal model parameter estimation method using mean velocity as an extracted feature and the particle filter. In addition, an estimation methodology for the microscopic traffic situation is developed to set the initial condition of simulation for forecast in accordance with observation. The proposed methods are verified by simulation-based traffic flow. The simulation adopts the stochastic Nishinari-Fukui-Schadschneider model. The optimal model parameters are successfully derived from posterior probability distributions (PPDs) estimated from DFOS data. In contrast, those estimated from point sensors fail. The PPDs of model parameters also indicate that each parameter has different sensitivities to traffic flow. A traffic forecast up to 60 minutes later is carried out. Using optimal model parameters estimated from DFOS, the forecast error of mean velocity is approximately $\pm$10 km/h (percentage error is 18%). The error attains half of it when conventional point sensors are used. We conclude that DFOS is a powerful technique for traffic modeling and short-term forecast.