Research On Road Traffic Flow Speed Prediction Based On Deep Learning

With the rapid development of urban transportation,traffic data shows explosive growth,while road traffic efficiency is gradually reduced,and the capacity of urban road network is gradually difficult to meet traffic demand,resulting in frequent urban traffic congestion.At the same time,big data and intelligent complex computing are also developing rapidly,and artificial intelligence methods can predict future traffic more accurately than statistical models,especially deep learning networks.After obtaining rich spatio-temporal data,deep learning methods can model and analyze road traffic and predict traffic conditions.Therefore,the main challenge of traffic prediction is how to obtain spatio-temporal characteristics from the complex spatio-temporal dimensions of massive traffic data and dynamically model them,while also considering various factors affected by the actual environment.Based on the deep learning method,this paper conducts short-term and medium-to long-term speed prediction research on traffic flow,and the main work is as follows:(1)Propose a short-term prediction model of traffic flow speed based on multiple influencing factorsIn order to improve the accuracy of traffic speed prediction,this paper predicts the traffic flow based on the regularity analysis of the actual traffic flow,and considers the actual environmental factors that affect the traffic state,a short-term prediction model MF-BiLSTM for traffic flow speed is proposed.This model use the real urban road traffic flow time series data set and meteorological data set to obtain road traffic speed and multiple characteristic variables related to traffic speed,including weather,temperature,traffic index,air quality index,etc.,and establish a characteristic matrix after data processing.And input the Bidirectional Long Short-term Memory network for training and validation,and verify the usability of the model in different environments.(2)Based on Attention Mechanism and spatiotemporal Graph Convolutional Network,propose a Medium and Long-term traffic speed prediction model.The current traffic flow prediction is ideal in the short-term,but the long-term traffic prediction effect has not been very good.There are two main problems of complex spatiotemporal correlation and error propagation.For the medium and long-term traffic flow prediction problem,based on the attention mechanism,combined with the spatiotemporal graph convolution network,a medium and long-term traffic speed prediction model TA-GCN is established.The model can follow the time series,synchronously obtain the spatial characteristics of traffic data,and dynamically pay attention to the information of each node,reducing the accumulation of errors under long-term forecasting tasks.In this paper,the MF-BiLSTM and TA-GCN models are experimentally evaluated using public datasets,and the effectiveness of the proposed model is verified.First,four prediction models of ARIMA,CNN,LSTM and BiLSTM are selected to compare with MF-BiLSTM.The experimental results show that the prediction accuracy of MF-BiLSTM is 2.63%,1.95%,2.87% and 0.89% higher than the above four comparison models respectively.At the same time,the dataset is divided according to rainfall degree and air quality index,and the prediction accuracy of each model is compared experimentally under different rainfall degree and air pollution degree.Compared with other models,the prediction accuracy of MF-BiLSTM is more stable.Secondly,the real road time series speed data set is used to conduct a medium and long-term road speed prediction experiment,and three models of ARIMA,GCN,LSTM-GCN,SAtt+Tatt and TA-GCN are selected for comparison and verification.The original data interval is 15 min,and the future window of the prediction experiment is selected 30 min,45min,60 min,the prediction accuracy of TA-GCN is better than other models,and good prediction results are achieved.