Research On Urban Traffic Flow Forecasting Based On Vector Autoregression

Urban traffic is one of the people’s livelihood issues that people are most concerned about,and the city is jammed with a large number of vehicles to bring great pressure to the urban traffic.According to the 2021 China Urban Traffic Report,Beijing’s average commuting time has reached 47.6 minutes,making it one of the most crowded cities in the world.Therefore,traffic flow state prediction has become a hot topic today.Traffic data is a kind of classical spatio-temporal data.It usually has the characteristics of large-scale,high dimension,many lost data,strong spatiotemporal dependence and so on.If people can know the traffic flow status on the road in advance,it can help people greatly reduce travel time.Existing traffic flow prediction strategies have some limitations that they can not effectively deal with the problem of missing data in modern spatio-temporal data.Some traditional methods using convolution to extract time features are easy to lose long-term features,so it is difficult to make long-term prediction.In order to overcome the problems existing in the current algorithm,the main research results are as follows:(1)Aiming at many data missing problems in high-dimensional traffic data,this paper proposes an Incremental Bayesian Vector Auto-regressive Factorization(IBVARF)framework.It is used to model multidimensional time series in specific spatio-temporal data in the presence of missing data.The model combines matrix/tensor decomposition and vector auto-regressive model to obtain a probabilistic graphical model.Through a specific vector auto-regressive process,the algorithm can better identify the dependence patterns between different time factors.Finally,an incremental learning model is adopted to reduce the burden of training.Graphical models allow us to make probabilistic predictions efficiently and handle cases of extreme data corruption more efficiently.(2)Considering that urban traffic data is not only random in the short term,but also periodic(morning/evening peaks)in the long term,this paper proposes a Switching Vector Auto-regressive Factorization(SVARF)algorithm.This is a probability generation model,which is used for spatio-temporal data such as urban traffic.SVARF model extends Na?ve Bayes switched linear dynamical system and combines matrix factor decomposition technology.It uses nonparametric Bayesian method for learning.SVARF algorithm parameterizes the weight of time factor into Markov priors with discrete potential state,which can learn nonlinear time dynamics,so that it can predict in the traffic field with potential nonlinear time state transition.The SVARF algorithm can reveal repetitive patterns in the data,expand the scope of prediction,and improve the accuracy of prediction.For the above two models,several experiments are designed to verify the validity of the IBVARF model,which can accurately interpolate and predict short-term traffic data on traffic datasets with data loss rates of more than 50%.The SVARF model improves the ability to cope with long-term recurrence and short-term unexpectedness of traffic data,successfully captures the periodic patterns of day and week dimensions of traffic data,and performs robust long-term predictions.The experimental results show that the algorithm presented in this paper performs better than several recent benchmark algorithms in data interpolation and long-term prediction.