Short-term traffic flow prediction in smart multimedia system for Internet of Vehicles based on deep belief network

Highlights

• Application of deep learning in short-term traffic flow prediction.

• Proposing a novel RBM structure to adapt deep learning in prediction.

• Multimedia processing and communication technologies in IoVs.

Abstract

In the multimedia system for Internet of Vehicles (IoVs), accurate traffic flow information processing and feedback can give drivers guidance. In traditional information processing for IoVs, few researches deal with traffic flow information processing by deep learning. Specially, most of the existing prediction technologies adopt shallow neural network, and their models for chaotic time series are prone to be restricted by multiple parameters. Over the last few years, the dawning of the big data era creates opportunities for the intelligent traffic control and management. In this paper, we take Restricted Boltzmann Machine (RBM) as the method for traffic flow prediction, which is a typical algorithm based on deep learning architecture. Considering traffic big data aggregation in IoVs, multimedia technologies provide enough real sample data for model training. RBM constructs the long-term model of polymorphic for chaotic time series, using phase space reconstruction to recognize the data. To the best of our knowledge, it is the first time apply RBM model to short-term traffic flow prediction, which can improve the performance of multimedia system in IoVs. Moreover, experimental results show that the proposed method has superior performance than traditional shallow neural network prediction methods.

Introduction

The Internet of Vehicles has subverted the existing traffic mode, becoming a new intelligent transportation mode to promote the further development of Intelligent Transportation System (ITS) [1]. The Internet of Vehicles makes the vehicle no longer a single mobile individual [2]. Through multimedia system, a powerful information network is formed. Real-time and accurate short-term traffic flow prediction is the key to realize smart multimedia system [3], [4], which is also important for quality of the whole Internet of Vehicles. The smart multimedia system has higher requirements for real-time performance. Currently, the maximum cycle of traffic control about three minutes, and the cycle of traffic guidance is generally 5 min. Therefore, how to predict traffic flow accurately in 5 min is very important. However, the traffic flow system has the apparent chaotic characteristics, and the short-term traffic flow data sampling is a typical chaotic time series. The basic idea of prediction chaotic time series is to construct a nonlinear mapping to approximate the system, and this nonlinear mapping is the prediction model to be established [5]. Therefore, traffic flow data preprocessing and chaotic characteristics reconstruction are the basis to improve the accuracy and reliability of the smart city short-term traffic flow prediction.

Computational traffic flow prediction technologies can be classified into the following five categories: (1) Method based on time series prediction, such as index analysis method, constituent factor analysis method, additive model, multiplication model and so on [6], [7], [8]. (2) Method based on adaptive filtering algorithm, such as least mean square algorithm (LMS), transform domain adaptive filtering algorithm, affine projection algorithm, conjugate gradient algorithm, adaptive filtering algorithm based on subband decomposition, adaptive filtering algorithm based on $QA$ decomposition and so on [9], [10]. (3) Method based on nonlinear network system, such as logarithmic curve equation, inverse function curve equation, cubic curve equation, compound curve equation, power function curve equation and so on [11], [12], [13]. (4) Method based on regression analysis, such as linear regression, logistic regression, polynomial regression, stepwise regression and so on [14], [15]. (5) Method based on chaos theory, such as OGY method, continuous feedback control method (external force feedback control method and delay feedback control method), adaptive control method and intelligent control method (neural network and fuzzy control), etc [16], [17]. Although scholars have proposed a lot of traffic flow prediction methods, these methods almost belong to the category of shallow neural networks. The shortcomings of shallow neural networks are strictly limited to weights and thresholds, and the convergence rate is relatively slow and the learning rate is relatively fixed. This inspires to rethink the chaotic time series prediction model with the deep learning architecture.

As an emerging field of data mining, deep learning is a learning algorithm that simulates the data model by imitating the multi-layer perceptual structure of the human brain [18]. Deep learning reflects the empirical performance in the processing of images, text, voice and other unstructured data, etc. Restricted Boltzmann Machine (RBM) is a typical algorithm of deep learning architecture [19]. Different from the traditional neural network models, RBM joins the feature learning part on the basis of the original multi-layer neural network. The feature learning part is to imitate the human brain to deal with the data signal classification. The specific operation is to increase the partial connection of the convolution layer and dimension layer, in front of the original fully connected network layer [20]. In simple terms, the traditional shallow neural network projections step is from the feature mapping to the value, and the characters are artificially selected. RBM projections step is from the signal to the feature and then to the value. The data characteristics are freely chosen by the network [21], [22].

In this paper, we propose a novel model based on Restricted Boltzmann Machine (RBM), which is a typical algorithm of deep learning architecture based on neural networks. RBM develops a long-term model of polymorphic prediction for traffic flow, which considers the chaotic time series. The mutual information method is used to select the delay time and the virtual neighbor method is used to select the embedded dimension. The traffic data is reconstructed, which carries on the phase space reconstruction to initial data. By calculating the saturated correlation dimension and the maximum Lyapunov exponent, the data chaotic characteristics are judged [23], [24]. Data with chaotic characteristics is entered into the proposed RBM framework for training. In addition, experimental results illustrate that the proposed approach has superior performance than traditional shallow neural network prediction methods. Specifically, the following three major contributions can be summarized.

(1) We apply the deep learning architecture to the smart city short-term traffic flow prediction. To the best of our knowledge, it is the first time to apply RBM model to short-term traffic flow prediction in Internet of Vehicles. Prior to this, shallow neural network has made a lot of prediction experiments in the field of short-term traffic flow, but the prediction performance fluctuates obviously, due to the impact of incremental unlabeled sample data. The convergence and robustness of deep learning occupy a greater superiority.

(2) We construct a novel RBM structure based on deep learning theory. Different from general network structure, RBM structure is a bipartite graph, and all nodes are random binary variable nodes, which each layer of nodes are not linked. The input data layer is a visual layer, and the hidden layer is another expression of the visual layer. The hidden layer is a full probability distribution that satisfies the Boltzmann distribution, and the hidden layer can be used as a feature of the visual layer input data. Prediction performance to the short-term traffic flow with chaotic characteristics has been improved intrinsically.

(3) We propose the multimedia processing and communication technologies to IoVs. We use the smart multimedia system to collect and analyze the chaotic time series, and we introduce the false adjacent point method to determine the optimal embedded dimension. On this basis, the phase space reconstruction of short-term traffic flow sequence is carried out. The chaotic characteristics of the time series are determined by calculating the multimedia processing and communication technologies. In essence, we create the traffic flow preprocessing, which can less generalization error to improve the nonlinear fitting ability.

The rest in this paper is organized as follows. Section 2 will state the background and motivation of short-term traffic flow prediction. In Section 3, chaotic time series prediction technology based on the Restricted Boltzmann Machine (RBM) will be given. In Section 4, it will discuss the smart city traffic flow data preprocessing. In Section 5, the experimental results are discussed. At last, conclusion will be done in Section 6.