Research On ITS Short-term Traffic Flow Prediction Algorithm Based On Deep Learning

Traffic flow forecasting is one of the key technologies of Intelligent Transportation Systems(ITS).Traffic flow forecasting refers to the use of time series algorithms or intelligent calculation methods to predict future traffic information changes based on historical information and other influencing factors.Due to the external and internal interference of the traffic system,the traffic parameters have certain complexity and randomness.It is difficult for a single method or model to accurately describe the changes of traffic parameters.The deep learning method has strong nonlinear processing and feature learning.The ability has been widely used,and many combined analysis methods have appeared in recent research,which can combine the advantages of multiple algorithms.This paper takes the deep learning method as the basic prediction algorithm,and mainly studies three kinds of traffic flow prediction methods: several basic time series analysis methods,time-space prediction algorithm based on convolutional network(CNN),and empirical mode decomposition(EMD).And space-time prediction algorithms for long-and short-term memory networks(LSTM).The experiment gives the prediction curve analysis of each algorithm,and calculates the error indicators to measure the prediction performance of the model and compares them.The following conclusions can be drawn: When the data is the same,the traditional time series prediction analysis method is the largest,time-space correlation in terms of model error index.Secondly,the CNN improved algorithm is the smallest,and the space-time EMD-LSTM algorithm is the smallest.In the complexity and training efficiency of the model,the traditional time series algorithm and the LSTM-based algorithm are faster and less complex than the space-time related CNN algorithm.Finally,in all simulation methods,the space-time EEMD-LSTM algorithm using noise-assisted analysis is optimal.The main results obtained in this paper are as follows:Firstly,several short-term traffic flow prediction methods including exponential smoothing algorithm,SVM and BP neural network algorithm are studied.Several basic time series prediction methods can roughly simulate the overall trend of traffic flow and complete at the same time.The collection and cleaning of traffic flow data;followed by the use of CNN’s advanced structure-intensive connection CNN,combined with CNN’s inherent ability to extract spatial features,proposed a densely connected convolutional neural network using two time feature scales for prediction,with spatio-temporal correlation The densely concatenated convolutional network fusion prediction with multiple time dimensions can reduce the measurement error and improve the prediction effect.Finally,the EMD-LSTM algorithm with multiple LSTM hierarchical prediction using multi-dimensional IMF components is used to significantly reduce the prediction error.The EEMD algorithm and LSTM of noise-assisted analysis constitute a combined prediction algorithm.Compared with all the prediction methods mentioned above,the space-time EEMD-LSTM algorithm has the smallest prediction error and the best prediction effect.