Traffic Flow Prediction Based On An Improved LSTM Network

In recent years,with the development of deep learning,the usage of deep learning methods to solve the problems in the transportation field has received wide attention,especially the superiority of long-short memory network(LSTM)in dealing with time series problems.However,the potential of deep learning methods in traffic flow prediction problems has not been fully explored from the depth of model architecture,the spatial scale of prediction regions,and the predictive power of spatio-temporal data.This paper proposes a deep bi-directional LSTM and LSTM neural network structure that considers the forward and backward dependence of time series data,and is used to predict the traffic flow of highway sections.Taking into account the potential reverse dependence of the time series,the bidirectional LSTM layer is used to obtain the correlation of spatial features and bidirectional time from historical data.In this study,road traffic flow data is expressed in matrix form,and the spatial correlation characteristics of road traffic flow are fully considered.Meanwhile,bi-directional LSTM network is used to correlate the potential context information of traffic flow historical data in time and to fully learn the features.This study adds a bidirectional LSTM network layer module to the LSTM network,making full use of the contextual relevance of time series to better model traffic flow prediction problems.In addition,the scalable model predicts traffic flow on highways and complex urban transportation networks.Comparisons with other prediction models show that,the proposed LSTM neural network with bi-directional LSTM is better than some other deep learning traffic flow prediction methods in terms of prediction accuracy.