Machine-learning-based vehicle delay prediction at signalized intersections

vehicle delay, machine learning, prediction models, data-driven models, delay prediction.

Garshasebi, Behnoush ; Benekohal ; Rahim F.

关键词: vehicle delay, machine learning, prediction models, data-driven models, delay prediction.;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/102430/GARSHASEBI-THESIS-2018.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Delay is one of the critical elements of signalized intersections performance measures. Field delay calculations are usually time-consuming and inefficient and dominantly rely on manual observations and post-calculations or complicated infrastructure configurations of detectors and vehicular sensors; making quick and reliable decision making difficult for practitioners. In this thesis, we implement the machine learning methods such as Linear Regression (LR), Support Vector Regression (SVR), and Random Forest (RF) to construct fast and straightforward models for vehicle delay prediction at signalized intersections. The models are trained by the use of available data sources such as signal timing plan and queue length per cycle. A thorough comparison between the performance of different machine learning models is also provided. To the best of our knowledge, this is the first study that implements machine-learning approaches to develop prediction models for vehicle delay at signalized intersections. Our study area consists of five signalized intersections along the Neil Street corridor (Major Street), Champaign, IL. Four of the crossing streets create typical four-legged intersections, and one of them create three-legged intersection; including one Major Street and five Minor Streets. We have observed that among the three models, LR and RF models achieved the best performances in predicting vehicle delay per cycle on the major streets when cycle length, green time and queue length are used as the predictors. LR predictions for the vehicle delay per cycle on the major street had an average error of 5.3 seconds, compared to the field vehicle delay per cycle and the R-squared of the LR model is 0.62; which shows that the model is able to explain 62% of the variation in the vehicle delay on the major street. Also, for the LR, the average accuracy (1-MRE) was 55%, 63% and 51% for the combined dataset, major street dataset, and minor streets dataset, respectively. On the other hand, RF predictions for the vehicle delay per cycle on the major street had an average error of 5.2 seconds, compared to the field vehicle delay per cycle. For the RF, the average accuracy (1-MRE) was 61%, 65% and 55% for the combined dataset, major street dataset, and minor streets dataset, respectively. Furthermore, on the major street, LR had errors less than 6 seconds in 87% of the predictions; and RF had errors less than 6 seconds in 80% of the predictions. This study shows that the effectiveness of the LR model for vehicle delay prediction on the major street is promising and manifested. Therefore, the straightforward LR predictive model proposed in this study, can be applied as an efficient tool to understand the current level of service at signalized intersections quickly. Rapid performance assessment further identifies the problems of the signal timing plan and enables the practitioners to make decisions on necessary remedial actions.