Open AccessEffect of Time Intervals on K-nearest Neighbors Model for Short-term Traffic Flow Prediction
Author(s) -
Lei Zhao,
Xiao Qin,
Wei Huang,
Xuanbing Zhu,
Yun Wei,
Jinde Cao,
Jianhua Guo
Publication year - 2019
Publication title -
promet
Language(s) - English
Resource type - Journals
eISSN - 1848-4069
pISSN - 0353-5320
DOI - 10.7307/ptt.v31i2.2811
Subject(s) - interval (graph theory) , traffic flow (computer networking) , reliability (semiconductor) , prediction interval , term (time) , nonparametric statistics , benchmark (surveying) , computer science , flow (mathematics) , time point , measure (data warehouse) , statistics , mathematics , data mining , geography , power (physics) , physics , geometry , computer security , geodesy , philosophy , combinatorics , quantum mechanics , aesthetics
The accuracy and reliability in predicting short-term traffic flow is important. The K-nearest neighbors (K-NN) approach has been widely used as a nonparametric model for traffic flow prediction. However, the reliability of the K-NN model results is unknown and the uncertainty of traffic flow point prediction needs to be quantified. To this end, we extended the K-NN approach by constructing the prediction interval associated with the point prediction. Recognizing the stochastic nature of traffic, time interval used to measure traffic flow rate is remarkably influential. In this paper, extensive tests have also been conducted after aggregating real traffic flow data into time intervals, ranging from 3 minutes to 30 minutes. The results show that the performance of traffic flow prediction can be improved when the time interval increases. More importantly, when the time interval is shorter than 10 minutes, K-NN can generate higher accuracy of the point prediction than the selected benchmark model. This finding suggests the K-NN model may be more appropriate for traffic flow point and interval prediction at a shorter time interval.