Research on Urban Road Mean Speed Prediction Method Based on LSTM-CNN Model

Abstract

Urban road mean travel speed prediction is an important component of Intelligent Transportation System (ITS). The mean road travel speeds are affected by the factors such as the time-varying traffic demand and the weather with fluctuation at the microscopic level. How to measure the influences of these factors and quantify them is a problem that needs to be solved. Hence, a combined model based on a Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN) was proposed to predict the hourly road mean speed, and a multivariate dataset containing multidimensional features that impact the mean travel speed is also used. While the LSTM is employed to learn the temporal dependence of the road mean speed sequence, the CNN is used to capture the spatial influences arising from the cross-sectional data. For improving the prediction accuracy, the Median Absolute Deviation (MAD) method was used to deal with the outliers representing the abnormal traffic condition among the cab data in city of Xi’an. The results of the comparative experiments show that a more accurate prediction of the hourly road mean speed can be achieved after removing the outliers and using multivariate dataset, and the proposed model outperforms the LSTM model in prediction accuracy measured by the R-Square, the Mean Absolute Error (MAE), and the Mean Absolute Percentage Error (MAPE), and the ConvLSTM model which has a spatio-temporal feature extraction capability.