부산대학교 도서관

Prediction of vehicle speed profiles is vital to many transportation and vehicular applications. However, accurately predicting these profiles remains challenging due to the complex and uncertain nature of the factors influencing driver behavior. This research presents a novel data-driven model that leverages a deep two-stage long short-term memory (LSTM) architecture to effectively capture the relationship between vehicle speed and macroscopic road attributes. Our model integrates road features, obtainable from various online map services, and average speeds as inputs and generates naturalistic speed profiles for a given route. The ultimate goal of this study is to integrate the proposed model with a microscopic emission model and incorporate it downstream of a mesoscopic traffic model. This integration enables the generation of high-resolution spatiotemporal maps of traffic emissions. The proposed model is trained and evaluated on a large dataset, including various driving records. The results demonstrate its ability to accurately generate realistic driving patterns while reproducing fuel consumption and emissions levels similar to those of real-world profiles.