基于LSTM和灰色模型集成的短期交通流预测

发布时间：2018-03-16 12:41

本文选题：短期交通流预测　切入点：深度学习　出处：《南京邮电大学》2017年硕士论文　论文类型：学位论文

【摘要】：道路交通系统是一个国家国民经济发展的基础,建造一个合理高效的道路交通系统是至关重要的。随着人们出行需求的增多,大众和企业对道路交通系统的便捷度要求越来越高。为了解决道路拥堵的状况,我们着力于研究交通流预测技术,期望能获取未来短期时间内精准的车流量数据,以实现车辆分流、交通诱导、道路规划和交通设施合理分布等目的。本文主要的研究对象是交通流数据,研究目标是精确的预测某选定路段未来一天以内的交通流量,研究内容是交通流数据预处理、交通流预测模型搭建和预测效果检验,设计了预测精度较高的交通流预测算法。本文的主要研究内容和研究成果如下:(1)本文首先对交通流数据的参数、特征和影响因素进行分析,选取时间和车流量作为本文的研究参数。使用EViews数据分析器获取数据的季节性和趋势性特征,为选取合适的非线性预测模型做铺垫。然后对交通流数据进行预处理。使用SPSS数据分析器调整数据顺序、添加空缺值并改正非常规值;再对数据进行小波软阈值去噪,去噪过程包括小波分解、软阈值去噪和小波重构,使用matlab代码实现并获得去噪后的数据表和数据图。(2)建立交通流数据的LSTM模型和GM模型。LSTM模型使用keras框架和python代码编写。将预处理后的部分数据输入进搭建好的LSTM网络,LSTM通过学习数据的特征确定网络参数和权值,并输出未来一天的交通流数据。LSTM模型的预测效果较好,但是模型训练需要的数据量较大。GM属于灰色模型,我们采取10个数据一个模型,不断改变模型参数,构造动态灰色模型。GM模型的预测效果不如LSTM模型的预测效果好,但是预测所需的数据量较少且实时性强。(3)将LSTM模型和GM模型使用动态权值w集成。针对单个模型预测法在应对突发状况时容易遗漏和忽视,导致预测精度降低,采用两种预测模型集成的方式对交通流预测进行研究。集成方式为加权组合,权值w利用关联系数确定,权值的动态步调与GM的建模步调保持一致。集成模型的预测结果显示,其预测精确度比两个模型单独预测的精确度高。
[Abstract]:The road traffic system is the foundation of a country's national economic development. It is very important to build a reasonable and efficient road traffic system. In order to solve the problem of road congestion, we are working on traffic flow forecasting technology to get accurate traffic data in the short term in the future. In order to realize the purpose of vehicle shunt, traffic guidance, road planning and reasonable distribution of traffic facilities, the main research object of this paper is traffic flow data, the research goal is to accurately predict the traffic flow of a selected section of the road within one day in the future. The content of the research is traffic flow data preprocessing, traffic flow forecasting model building and forecasting effect testing. A traffic flow forecasting algorithm with high precision is designed. The main contents and results of this paper are as follows: (1) in this paper, the parameters, characteristics and influencing factors of traffic flow data are analyzed. Time and traffic flow are selected as the parameters of this paper. The seasonal and trend characteristics of the data are obtained by using the EViews data analyzer. In order to select the suitable nonlinear prediction model, the traffic flow data is preprocessed. The SPSS data analyzer is used to adjust the data order, add the vacant value and correct the unconventional value, and then the wavelet soft threshold is used to de-noise the data. The denoising process includes wavelet decomposition, soft threshold denoising and wavelet reconstruction. The LSTM model of traffic flow data and GM model. LSTM model are written using keras framework and python code. The pre-processed part of the data is input into the constructed LSTM. The LSTM determines the network parameters and weights by learning the characteristics of the data. And output the traffic flow data. LSTM model in the next day has good prediction effect, but the model training needs a large amount of data. GM belongs to the grey model. We take 10 data and one model, and constantly change the model parameters. The prediction effect of dynamic grey model. GM model is not as good as that of LSTM model. But the amount of data needed for prediction is less and real-time. 3) the LSTM model and GM model are integrated with dynamic weight w. The prediction method of single model is easy to be omitted and ignored when dealing with sudden situation, which leads to the decrease of prediction accuracy. Traffic flow forecasting is studied by two integrated forecasting models. The integration method is a weighted combination, the weight w is determined by the correlation coefficient, the dynamic step of the weight value is consistent with the GM modeling step, and the prediction results of the integrated model show that, Its prediction accuracy is higher than that of the two models alone.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.14

【参考文献】