The Research On Prediction Model And Prediction Method Of Short-term Traffic Flow

In this paper, we do a lot of research on prediction model and prediction method of short-term traffic flow. We propose a new multi-link model for traffic flow forecasting basing on Neural Network, and apply Gaussian Process Regression in traffic flow forecasting for the first time. Furthermore, we construct a more efficient multi-link model using Graphical Lasso.Short-term traffic flow forecasting is always a hot topic in Intelligent Transportation System. Traditional prediction models are normally single-link prediction models, that is, they predict the future traffic flow on a link using only the historical flows on the same link but not take the flows on the adjacent links into account. In fact, each link is related to the other links in the whole transport system, especially the adjacent links. In this paper, we put forward a multi-link prediction model which takes the relations between adjacent links into account. Through sufficient experiments, we verified the superiority of multi-link model.Neural Network is well known in machine learning area. Especially, BP (Back-Propogation) network is widely used in many areas such as visual scene analysis, speech recognition etc. Due to the excellent ability in handing complex problems, and the characteristic of self-learning, self-organization and self-adaption, Neural Networks always do well in machine learning. In this paper, basing on Neural Network, we make experimental comparison between single-link and multi-link. Moreover, combing with single-task and multi-task learning, we construct four models in all. Through the global and local comparison of the result of all the corresponding experiments, we achieve a comprehensive understanding of Neural Networks used in traffic flow forecasting.Gaussian process regression is a classic regression approach basing on Bayesian theory. Due to the characteristics of tractability, interpretability and few-parameters, it is widely studied in machine learning. Gaussian process is a generalization of Gaussian probability distribution. In Gaussian probability distribution, the random variables are scalars or vectors (multivariate distribution case), while the random variables are functions in Gaussian process. Inference in Gaussian process is also taken in function space. Gaussian process regression prediction algorithm outputs a posterior distribution on the target of the training set. This paper does abundant theoretical analysis of Gaussian process regression and finally applies it in actual traffic flow forecasting.Graphical Lasso is an approach of constructing sparse graphical model. Through adding a L1 regularization term to the inverse covariance matrix, Graphical Lasso aims to make the matrix as sparse as possible. Each row or column of the matrix represents a node in the graphical model, which also corresponds to a variable of the data set. If there is a zero component in the inverse covariance matrix, then there is no link between the two corresponding nodes in the graphical model. In other words, the two variables are independent given other variables. In this paper, we give theoretical analysis of Graphical Lasso, and combine the sparse model with actual prediction models. Experimental results show Graphical Lasso is an efficient approach in extracting information and it brings a further improvement of multi-link model for short-term traffic flow forecasting.