ZHANG, Ce (University of Waterloo), Muresan, Matthew (Miovision Technologies Inc.), PAN, GUANGYUAN (Linyi University), Fu, Liping (University of Waterloo), Lu, Zhengyang (University of Waterloo)

Temporal Fusion Transformer for Real-Time Intersection Turning Movement Flow Forecasting Incorporating Exogenous Factors

Scheduled for presentation during the Poster Session "Transformer networks" (WeBT13), Wednesday, September 25, 2024, 14:30−16:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Abstract— Real-time intersection turning movement flow (TMF) forecasting is vital for traffic management systems to optimize traffic flow and signal timing. Recent studies have introduced advanced machine learning models for TMF forecasting to capture intricate patterns and relationships present in the data. However, most models rely solely on past observed traffic flow data, neglecting the influence of exogenous factors. This paper addresses these limitations by investigating the potential exogenous factors that could affect TMF and quantifying their impact on forecasting accuracy. To achieve this, a case study is conducted that involves real-world traffic data from 76 intersections with 17 exogenous factors. These factors encompass static covariates (e.g., speed zone, road type), prior-known future time-dependent inputs (e.g., hour of day, temperature), and past observed TMF. Experimental results show that, by incorporating these factors, the TFT model achieved superior performance with lower forecasting errors across various data subsets and whole dataset compared to other models that solely rely on time series data. The modeling results also reveal that speed zone and road category are the most influential static covariates. Among the time-dependent covariates, hour of day and temperature have the strongest influence.