Traffic Speed Prediction Under Non-Recurrent Congestion: Based on LSTM Method and BeiDou Navigation Satellite System Data

摘要

The full utilization of Location-Based Vehicle Sensor Data (LB-VSD) can improve the efficiency of traffic control and management. Currently, LB-VSD is widely applied to the prediction of traffic speed. Like the GPS system, BeiDou satellite navigation system (BDS) can collect LB-VSD. In China, the key operation vehicles on the expressway are equipped with BDS to monitor the travel path. This provides a basis for predicting the traffic speed on expressway accurately. In this paper, considering the abnormal data collected by BDS, the screening and processing rules are made, and then the traffic speed sequence is extracted. Considering the data-missing problem caused by equipment failure or abnormal data elimination and the data sparse problem caused by small size of sample, a filling method based on trend-historical data is proposed. Traffic flow evolution is a complex process. Sudden accidents or bad weather can cause a sudden change in traffic flow and non-recurrent traffic congestion. The prediction accuracy of traditional machine learning methods is low when non-recurrent congestion occurred. In order to solve this problem, this paper adopts a deep learning model-Long Short-Term Memory (LSTM) to predict the traffic speed. Moreover, three-regime algorithm is used while building the prediction model. The prediction method is compared with Support Vector Regression (SVR) method. The results show that the prediction accuracy of the proposed method is higher than that of SVR algorithm, and the robustness is better in the case of non-recurrent traffic congestion.