Research On Traffic Flow Prediction Based On Long Short-Term Memory Neural Network

Efficient and accurate traffic flow prediction is helpful to alleviate traffic congestion and improve traffic efficiency.In order to fully extract traffic features and improve the accuracy of traffic flow prediction,this paper utilizes the traffic flow data of urban road network to put forward two combined models of traffic flow prediction,which from the perspectives of reconstruction the input of model and traffic flow decomposition.The specific research contents are as follows:Based on the historical traffic flow data of a single intersection of urban roads,we propose a traffic flow prediction model based on similarity measurement,i.e.FLW model.Firstly,use Dynamic Time Warping(DTW)algorithm to measure the similarity in traffic time series,utilize Fuzzy C Means(FCM)clustering algorithm to classify traffic change patterns,and calculate the membership degree of each data point.Secondly,construct Long Short-Term Memory(LSTM)neural network model for each traffic change pattern,and predict the traffic flow in the future.Then,use the membership matrix to weight the predicted values of sub-models to get the final prediction results.Finally,we carry out experiments to compare the single model with the combined model.The results show that the proposed method has smaller prediction error and less residual fluctuation.The proposed method can identify the similar features accurately in the data,reduce the prediction error,improve the prediction accuracy,and provide a theoretical basis for urban road traffic flow prediction.Based on the traffic flow data of urban road network,we propose a traffic flow prediction model based on residual processing,i.e.LSTM-ARIMA model.Firstly,use Pearson correlation coefficient to analyze the correlation among all detection positions,and establish LSTM prediction model to predict the traffic flow of all detection positions synchronously in the road network.Secondly,carry out the stationary test and white noise test for residual series of the LSTM model.If the residual series are tested as non-white noise,we use the Autoregressive Integrated Moving Average(ARIMA)model to analyze and predict the residual series.Then,the predicted values of LSTM model and ARIMA model are fused to optimize the prediction results of LSTM model and improve the prediction accuracy.Finally,aggregate the data into 5 min,10 min,and15 min,respectively,and predict the traffic flow.In order to verify the necessity of residual processing technology and the excellence of the proposed model,we compare the prediction errors of the proposed model and the baseline models.The results show that compared with the baseline models,the predicted values of the proposed method are closer to the observed values,and the prediction accuracy is higher.The proposed method can improve the prediction performance by extracting the traffic information fully in residual series,and it has certain practicability in the application of urban road network traffic flow prediction.Based on the traffic flow data of urban road network,this paper proposes two combined prediction models,which can predict the traffic flow of a single intersection and the road network.The experimental results show that compared with the baseline models,the proposed method has a smaller prediction error,the prediction effect is better,which has a certain reference significance in the traffic flow prediction of urban road.