Research On Short-term Traffic Flow Prediction Based On LSTM

Short-term traffic flow prediction can help urban traffic management departments to complete the induction and control of traffic,reduce the degree of urban congestion,and improve the city’s operational efficiency.Improving the prediction effect is one of the key points of the development of this technology.In this paper,the section of the intersection area is selected as the prediction section.The prediction method selects the Long Short Term Memory Network(LSTM)with long-term "memory" capability in deep learning theory.The main research contents are as follows:1.First,determine the structure of the input data of the model,and analyze the spatio-temporal characteristics of the traffic flow for this purpose: in the time dimension,the time series analysis and study of the predicted road segments are conducted to determine the top 5 past ones that have a significant correlation with the current moment.Historical traffic data at the moment;on the spatial dimension,analyze the spatial correlation between the predicted road segment and the adjacent road segment,and confirm the adjacent road segment that has a significant correlation with the predicted road segment,thereby constructing the spatio-temporal correlation matrix of traffic flow data;2.Next,make a mathematical description of the traffic flow prediction problem,and use the historical traffic flow data before the current time of the predicted road section and the adjacent road section that is significantly related to its existence to predict the current traffic flow of the predicted road section;on the original LSTM network Improvements were made,using the input data of the convolution and pooling processing models,to construct a short-term traffic flow prediction model based on LSTM in this paper,and perform parameter optimization processing on the model.3.Finally,design an empirical program,conduct example tests from three aspects: the prediction accuracy of the model,the influence of adjacent road sections on the prediction accuracy of the model,and the long-term "memory" ability of the model,and draw conclusions:(1)The prediction accuracy of the model in this paper Nearly 95%,compared with LSTM single-point prediction model,BP neural network model(Back Propagation,BP)and support vector machine model(Support Vector Machine,SVM)have better prediction effect.At the same time,the model has higher prediction accuracy for other road segments,indicating that the model is portable.(2)Changing the number of adjacent road sections of the input model will affect the prediction accuracy of the model in this paper.Specifically,with the increase in the number of effective road sections introduced,the overall prediction accuracy of the model is on the rise.(3)In the case that the length of the input historical data takes different values,compared with the other models mentioned above,the prediction accuracy of the model in this paper changes little,which verifies the long-term "memory" ability of the LSTM network.