Multi Long Short-Term Memory Models And Spatio-Temporal Correlation Based Short Term Traffic Flow Prediction

With cars gradually becoming the main means of transportation,the problem of urban traffic congestion is becoming more and more serious.In order to improve traffic conditions,the role of Intelligent Transportation Systems(ITS)has received increasing attention.In this system,accurate and real-time short-term traffic flow prediction can be used as an important basis for traffic guidance.Therefore,it has become a research hotspot to find a suitable method to predict the change of traffic flow after a short period of time.Over the years,many methods have been tried to be applied in the field of traffic flow prediction,most of which are based on a single model and predict the change of traffic flow at the next moment of an observation node according to the historical traffic flow observation value of only that certain observation node.However,traffic flow is a kind of data with nonlinearity,randomness and spatial correlation,containing multiple components,which makes the prediction effect of single model and the method that ignores the spatio-temporal information bottleneck.In the view of these two problems,this thesis proposes two forecasting methods of short-term traffic flow based on Multi Long Short Term Memory Recurrent Neural Networks Models and Spatial-Temporal Correlation Long Short Term Memory Recurrent Neural Networks.The basic idea of Multi-model LSTM is to classify the traffic flow data properly and then establish the corresponding prediction sub-model according to the data of different categories.By extracting a feature from each traffic flow data sample for K-Means clustering,so that the samples can be adaptive divided into two categories with different changing trends,and then the corresponding LSTM sub-model can be trained for these two types of data samples.At the same time,it is necessary to use these two kinds of data to train a classifier composed of K-Nearest Neighbor algorithm,so as to meet the demand of real-time online prediction of traffic flow in practical application scenarios.Spatio-temporal correlation LSTM makes use of the spatial correlation relation between the node to be predicted and several observation nodes of its upstream and downstream,and takes the traffic flow data observed at all nodes as the input samples.At each moment,the correlation coefficient between the node to be predicted and other observation nodes should be required,and they should be taken as the weight coefficients of the LSTM input layer.This model training method can not only adaptively adjust the influence weights of different observation nodes on the node to be predicted in real time,but also enable us to observe which nodes in the traffic network are more important.In this thesis,the traffic flow collected by Caltrans Performance Measurement System is taken as experimental data.The mean absolute percentage error and root mean square error are used as evaluation criteria to compare the prediction accuracy of various mainstream prediction models and the two models proposed in this thesis.The experimental results show that the two proposed prediction models have better prediction effectiveness,and the fusion of the two models can further improve the prediction performance.