Research On Grid-based Short-term Traffic State Forecasting

According to statistics from related departments, the development of urban transport, especially the growth of private cars in recent years, has increased the issue between vehicles, pedestrians and roads increase. These result in many serious problems such as road congestion, frequent accidents and environmental pollution, which have become a bottleneck of the city's sustainable development. The conflict that transportation capacity cannot satisfy economic development and people's needs becomes more severe. The phenomenon that transportation capacity cannot satisfy economic development and people's needs will continue to rise and the conflicts will becomes more severe. A simple increase in road infrastructure is not a good solution to the transport, so Intelligent Transportation System (ITS) comes into being. As an important topic in the research field of ITS, Short-term traffic state forecast is the key question for the traffic guidance system.Short-term traffic State forecast means real-time forecasts for the next time t +⊿t (⊿t less than 15min), and even some time after it, based on the information collection at the time point t. Decision-making, according to this forecast, can guide and control the traffic flow migration, coordinate and balance the traffic flow, which plays an important role in alleviating traffic jams. With regard to real-time performance requirements, the typical cycle of traffic control and guidance is usually to be 5min, so it is critical that accurate forecasts of traffic state are made within 5 minutes. Traditional forecast model and nonlinear dynamics theories with characteristics of the short-term traffic state are analyzed in the paper and improvements the nonparametric regression model. Using the free resources of network builds a platform for the forecast system and information publishing, which solve the conflict between the prediction accuracy and efficiency ,ensuring the efficiency and accuracy of the method.Since the 1960's and 1970's, domestic and foreign scholars have put forward more than 30 species of short-term traffic state forecasting methods. According to principles, they can be divided into three categories: models based on the analysis of mathematical, models based on the simulation of traffic and knowledge-based intelligent models. The first kind of model usually uses the analytical mathematical model to describe the state variable traffic trends, which uses mathematical statistics and the assumption that the future traffic data and historical traffic data have the same characteristics. These models are simple but its accuracy is poor, and is especially not suitable for the sections with larger changes in traffic state. The second models are based on the simulation of traffic, including the model based on traffic simulation and the model based on dynamic traffic distribution. These models realize the traffic simulation -- an important tool for traffic analysis, but with non-real-time characteristics. Generally, they are used for the evaluation of existing models. The knowledge-based intelligent models, including non-parametric regressive forecast model, neural network model and Chaos model. These models points to the view of non-linear and unpredictable characteristics in the transportation systems, and fits it in with the real traffic state data, then forecasts traffic state for the next period of time. Such methods of forecasting are better, but it requires a large amount of calculation. Transport system is essentially a giant complex system, which is full of nonlinear interactions. So we choose the integrated model based on non-parametric regressive algorithm to research, which has advantages in principle. The main idea of the nonparametric regression mode is try to find the nearest neighbors in historical sample database, and then forecasts the state for the next time. Nonparametric regression model does not make historical data smoothing. In a time of incident, the nonparametric regression model can give better results than the parametric regression model.To ensure the efficiency and accuracy of the method, we research the principle of general nonparametric regression model, and point to four key steps to improve it.1. The improvement for the accuracyFirst, using principal component analysis gets the integral parts of the traffic state vector, and then proposes the traffic state vector structure based on the hierarchical object, which can not only clearly expresses the complex traffic state vector but also plays a important role in data compression and reduce the size of traffic state sample database.Second, in order to reduce redundancy of the data, and access to both typical and completeness of the historical sample database, this paper presents a clustering method of changing k center-density, which ensuring the forecast accuracy.Third, in order to obtain the true nearest neighbors, we present a new similarity measuring function of Data Related Weighted (DRW). According to the impact factors of the various dimensions components of the traffic state vector and the relative error of the data obtains the true nearest neighbors to reducing the prediction errors. In general non-parametric method, the value of k is changing with the size of historical database, which cannot satisfy with the need of promotion. So the paper raises a new search algorithm of searching radius R based changing k neighbor search algorithm. By defining the threshold of the distance between the traffic state vectors, gains all the nearest neighbor of the vector, which can raise the accuracy and general applicability of the forecast model. Last but not the least,in order to ensure the self-learning ability of the prediction model, the paper raise the algorithm with error feedback. By using the difference between the predicted result and measured data to adjust the value of R, the loss of neighbor is reduced and the accuracy is improved.The experimental results show that the improved non-parametric regression model for short-term traffic state forecast has higher prediction accuracy.2. The improvement for the efficiencyBy changing the traditional Hash-based optimization approach, the paper presents a optimization strategies based on the nested table of database, which can not only obtain higher query efficiency, but also prevent the hash conflicts and reduce the costs of database maintenance. The experimental results show that the query efficiency is basically the same between nested tables based structure and Hash-based organizational structure of database with a large number of data. But the efficiency of maintenance the nested tables based database has been greatly enhanced. In addition, by defining the max-similarity function in the database, by use of the database's own query optimization features to reduce the number of IO operations improve the operational efficiency of the program. The prediction model of nonparametric regression are needed to deal with a lot of sample data in a limited time, that results in a higher demand for computer processing power and calculation strength. No single device can afford it. And it will lead to high input costs, if a large number of hardware upgrades. So this article brings the combination technology of HA Cluster and Grid to make full use of idle computing resources in grid. By using the grid resources management mechanisms and job scheduling strategies, solves the efficiency of the prediction algorithm.By combining the hierarchical model and point to point model, the paper presents an intelligent method of self-governing domain (SGD) grid resource management with macro-control and quick response strategy. On the basis of SGD, we build the Multi-Hierarchy Self-Governing Domain (MHSGD) grid system, with the minimum price of migration. MHSGD divides the grid into multiple independent self-governing domains logically and minimize the distance between the brother nodes in the same domain. Thus ensuring the time of thus ensuring the grid status maintenance and job scheduling algorithm can be completed within the shortest possible time. In addition, this paper defines the computing power exponent (CPE) and the data migration cost (DMC), then establishes the task allocation model based on CPE and job scheduling model based on the minimize DMC, which can make full use of the resources in the grid, to reduce the waste of resources and to ensure load balancing.To make short-term traffic state prediction to better serve for the traffic management and traffic guidance systems, the paper presents the information release platform based on synchronous data cache to release the forecast results. It not only can meet with the large amount of concurrent accesses, but also avoid duplicated calculations of the same predict job, and reduce the burden on the grid system.At last, the paper builds experimental environment to verify the efficiency of the model. Using the large size of simulation data by vissim, the grid system of MHSGD can increase the efficiency of the improved non-parametric regression model for short-term traffic state forecast.