Deep Learning Based Traffic State Discrimination And Short-term Traffic Flow Prediction Methods

In recent years,with the continuous development of the economy,the rapid growth of car ownership has brought negative impacts such as traffic congestion and environmental pollution on urban roads.The emergence of intelligent transportation systems has played an important role in improving the efficiency of transportation operations.Reasonable traffic state discrimination and accurate traffic flow prediction provide important support for important modules such as traffic guidance and traffic planning in intelligent transportation systems,and are currently hot topics in the field of intelligent transportation systems.Aiming at the characteristics of spatio-temporal changes such as periodicity,similarity,and spatial connectivity of traffic flow,this paper takes the Los Angeles urban highway network as the research object,and conducts in-depth research from traffic state mining and short-term traffic flow prediction.Propose an improved traffic state discrimination and short-term traffic flow prediction method based on deep learning.This work can provide support for large-scale highway network management and control,and has great practical application value.Main tasks as follows:First,visually verify the three main characteristics of traffic flow changes,such as periodicity,similarity and spatial connectivity,to lay the foundation for traffic state discrimination and short-term traffic flow prediction.Secondly,considering the different trends of the traffic information sequence on weekdays and rest days,this paper constructs a classification sample pair,encodes the classification sample pair through a shared parameter LSTM network,and uses the characteristics of the similarity of the traffic state changes of different nodes of the road network.This coding method is extended to any length of traffic information sequence of any node in the road network;in order to mine more reasonable traffic state information,an improved K-Means clustering algorithm based on LSTM-encoded traffic information sequence is proposed and passed Experiments verify the effectiveness of the coding method and the convergence of the clustering method.Finally,for the short-term traffic flow prediction problem,combining the FM structure,attention mechanism and LSTM structure,a Cluster-Attention-LSTM prediction model is proposed to learn the periodicity and spatial connectivity information of traffic flow changes in both time and space.Learning different data change rules for different detector nodes effectively improves the accuracy of short-term traffic flow prediction;and based on the results of the improved clustering algorithm,on the premise of ensuring the prediction accuracy,the number of prediction model parameters is effectively compressed.Through experimental comparison and analysis,the prediction accuracy of the prediction model in the overall road network is verified,which shows the feasibility of the short-term traffic prediction model proposed in this paper.