Traffic Flow Prediction Model Based On Echo State Networks And Related Research

With the rapid development of economy and urbanization, some traffic problems caused by the increasing vehicle inventory, such as traffic congestion, traffic accidents and energy dissipation, are becoming one of the world’s biggest issues that hinder the urban development. As the most effective method for solving this problem, the Intelligent Transportation System (ITS) has been paid great attention in recent years. As one of the core part of ITS, the system of traffic control and guidance dependents on accurate prediction of traffic flow. Hence, realizing accurate and fast prediction of traffic flow has become one of the priorities of ITS development.However, when traditional models are used to predict single-step traffic flow time series with the sampling interval of two minutes, their error is often close to20%, and this low prediction accuracy is far from being able to meet the needs of rapid development of ITS. Moreover, because of the special application background of traffic flow prediction, we expect that the single-step prediction model used in practice should not only satisfy the requirement of speed and stability but also be feasible for multiple step prediction. These constraints make the research of traffic flow prediction trap into great difficulty. Hence, China’s Ministry of Science and Technology launched the973program Basic Science Research of Big City Traffic Congestion Bottleneck; in its core sub-topic Research of Intelligent Traffic Information Fusion and Integration, the traffic flow prediction is explicitly treated as its core content. We found that many models have achieved good results in the single-step prediction of traffic flow in recent years, but those models are over-focused on the optimization of their performance and seldom consider the physical characteristics of traffic flow, so it is difficult to break through the bottleneck of the low prediction accuracy, resulting in the difficulty and challenging of single-step and multi-step prediction of the traffic flow.As a new machine learning method, Echo State Network (ESN) proposed in the journal of Science in2004has the characteristic of distinct organizational structure and a strong short-term memory. So ESN can get a thousand times more precise than previous models to predict time series, especially the noise-free chaotic time series generated by deterministic systems such as Mackey-Glass and Lorenz sequence. Moreover, since ESN adopts the pseudo-inverse method to train its weight, it has both characteristics of fast training speed and high stability. Because of its various advantages, ESN are becoming a research focus in the area of machine learning and has been widely used in time series prediction in recent years. Because the real-world traffic flow belongs to a kind of time series with chaotic property, inspired by the classical chaotic time series prediction approaches, we introduce the ESN model into the prediction of traffic flow. However, we found that, if we directly applied the ESN to the prediction of the traffic flow, the prediction effect is not good due to the affection of the complex noise components in the traffic flow.In order to solve the problem of noise disturbance, based on the analysis of nonlinear dynamic of traffic flow, this paper integrates advanced signal processing methods and constructs several novel single-step or multi-step prediction models based on the ESN to predict short term traffic flow. Each of these models is able to improve the prediction accuracy, training speed and stability in certain degree. Based on the above research works, the main works and contributions can be summarized as follows:(1) Complex dynamics of traffic flow are analyzed from multiple perspectives. According to our research, we find that traffic flow is chaotic time series with strong noise. These noise components destruct the behavior of chaotic attractor of traffic flow and significantly reduce its predictability. In this paper, we qualitatively and quantitatively analyze the predictability of traffic flow in detail on the basis of dynamics analysis on traffic flow. We revealed the complex physical properties of traffic flow in essence. Those research methods and results also provide a solid theoretical foundation for the further construction of new prediction models.(2) ESN based single step and multiple step prediction models are proposed. As a novel prediction model, ESN is able to predict classical chaotic sequence in multi-step with a high precision. On the basis of dynamics analysis on traffic flow, this paper proposed traffic flow single step and multiple step prediction models based on ESN. Compared with the prediction results of several classic models, it is proved that the new models can improve the precision slightly and appear obvious advantages both in training speed and stability. After discussing some important parameters of the models, the optimal parameters of the ESN model for traffic flow prediction are assigned. In the meanwhile, this paper emphasizes the discussion of the performance of multi-step prediction while using the iterative method and direct method. Through analyzing their advantages and disadvantages, this paper revealed the source of the bottleneck of low prediction accuracy when using the traditional multi-step prediction models. This provides an important theoretical support for the following problem of multi-step prediction.(3) The multi-reservoir echo state network model based on multi-scale decomposition of wavelet domain (MESNMW) model is proposed to precisely predict traffic flow. This paper investigates the prediction accuracy problem that is not well solved in Part (2) due to the interference of noise component in traffic flow. We found that previous machine learning methods based prediction models focus on the studies of the generalization ability of models while ignoring the tremendous influence of noise components on the prediction accuracy. In order to avoid this influence, this paper proposed a novel MESNMW model from the perspective of signal processing. Using multi-scale decomposition method based on wavelet, this model is able to shield the noise components in traffic flow by treating them as the high-frequency components with less weight such that guarantees the main components, especially the low-frequency components, are chaotic sequences with high signal-to-noise. Combined with multi-reservoir ESN, its single-step prediction accuracy is nearly20times of that obtained by the simple ESN model. Besides, this paper also discusses the influences of some important parameters on the model performances to provide a theoretical basis for the proper choice of model parameters.(4) The multi-reservoir echo state network based on multi-scale of wavelet packet decomposition domain (MESNMWP) model is proposed to further improve the prediction precision of traffic flow. On the basis of MESNMW, this paper propose a novel MESNMWP model, which uses multi-scale decomposition approach based on wavelet packet to shield the noise components of traffic flow by treating them as high-frequency components with less weight, so that further guarantees the high signal-to-noise of the main components. The experimental results show that the prediction accuracy of MESNMWP single-step model can be increased to three times as that of MESNMW. In addition, this paper also adopts MESNMW and MESNMWP to complete both iterative and direct multi-step prediction of traffic flow. Since the iterative method can obtain more accurate prediction results, this study revises the one-sided traditional view "direct multi-step prediction method is more effective for real-world data". At the same time, it also provides a new way for multi-step prediction of other similar real-world data series.(5)The single-reservoir echo state network model based on multi-scale of wavelet decomposition domain (SESNMW) and single-reservoir echo state network model based on multi-scale of wavelet packet decomposition domain (SESNMWP) are proposed to achieve high prediction accuracy and efficient on traffic flow. Although the above MESNMW and MESNMWP have high prediction accuracy, their computational costs are usually very high. This will lead to the growth of training time, which greatly hinders the practical application of MESNMW and MESNMWP for traffic flow prediction. In order to ensure both the prediction accuracy and the computational efficiency, this paper proposed SESNMW and SESNMWP from signal smoothing perspective. Aiming at the conflict of the similarity and predictability of the SESNMW and SESNMWP, the multi-state threshold method is introduced into the SESNMW and SESNMWP to reduce noise influences. Compared with other traditional threshold method, this model can more effectively solve the conflict of the similarity and predictability, increasing the fitting degree of the traffic flow after noise reduction process and enhancing the prediction accuracy of single-reservoir ESN. This method can comprehensively assure the prediction accuracy and operating efficiency of SESNMWand SESNMWP.In summary, this paper focuses on exploring traffic flow prediction models. Firstly, based on the analysis of dynamics property of traffic flow, we propose several single and multi-steps models based on the ESN to improve the operating efficiency and stability of traffic flow prediction. Then, we establish MESNMWto implement high precise single and multi-step prediction of traffic flow. Based on MESNMW, we further create MESNMWPto improve the prediction accuracy. Finally, this paper proposes SESNMW and SESNMwp to satisfy the requirement of accuracy, speed and stability comprehensively for traffic flow prediction. The proposed traffic flow prediction models have good practicality, so they have important value of theoretical and practical application in the field of scientific research and engineering.