【作者】 罗向龙；

【导师】 张生瑞；

【作者基本信息】 长安大学 ， 交通运输规划与管理， 2021， 博士

【摘要】 随着社会经济的发展,现有的公路基础设施已经很难满足人们的出行需求。无论在高速公路和城市道路交通拥堵已经成为困扰交通管理部门的难题。信息技术、控制技术、网络技术等的飞速发展,产生的智能运输系统(Intelligent Transportation Systems,ITS)是全球公认的解决交通拥堵等问题的主要途径。近些年来,物联网、互联网+、大数据、人工智能等新技术的出现,更是给ITS注入了新的活力。如何利用最新的人工智能技术对交通状态做出预测,以便出行者选择合适的出行路线,从而减少出行时间、交通拥堵和环境污染成为ITS研究的热点。在对交通流预测的国内外研究现状分析的基础上,利用皮尔逊相关系数、相关函数定性与定量对交通流的时空相关性进行了分析。在时间特性上,分析了交通流工作日和非工作日以及连续相同工作日的相似性以及周期性,从时间角度为交通流预测模型数据集的选择奠定了理论基础。在空间特性上,对路网中不同路段交通流的相关性以及局部相似性进行了分析,结果表明路网中不同路段交通流的相关性与路段之间的距离有关,但相关性并不与路段的距离成反比,为交通流预测模型的构建奠定了基础。对交通数据异常产生原因,交通异常数据监测的肖维勒、狄克逊、莱依达方法进行分析的基础上,提出了一种局部莱依达准则的交通数据异常检测方法,并通过实验验证了提出方法在异常检测的可靠性和有效性;分析了交通数据缺失的原因并对缺失数据进行了分类,针对结构性交通数据缺失,提出了一种改进的基于矩阵低秩分解的结构性数据修复方法,并以实际的交通流量数据对提出方法的性能进行了测试和分析,表明了提出方法的有效性。以离散图信号处理的基本理论为基础,结合交通流的时空相关性,构造了一种基于数据和距离相关的路网邻接矩阵,并提出了一种基于图傅里叶变换的交通流时间序列和路网交通流数据分解方法,将交通流数据分解成基本的趋势项和随机波动项。充分考虑交通流的时空相关性,利用K近邻(K-nearest Neighbor,KNN)算法对路网中的不同路段的交通流数据进行筛选,在基于图傅里叶变换(Graph Fourier Transform,GFT)交通数据分解的基础上,结合深度学习模型中长短期记忆网络(Long Short-Term Memory,LSTM)处理时间序列的优势,提出了一种融合KNN、GFT和LSTM的组合交通流预测模型KNN-LSTM。实测的交通流数据测试结果表明,相比于传统的自回归滑动平均模型、机器学习中的经典模型支持向量机、浅层人工神经网络、深度信念网络和LSTM的深度学习模型,本文提出的方法具有更好的预测性能。针对现有关于大规模路网中所有路段交通流预测研究的不足,提出了一种基于K-means聚类分析和矩阵分解的路网交通流压缩预测方法。利用K-means聚类分析对路网中不同路段的交通流量进行聚类分析,以欧式距离作为判定标准将路网的路段划分为K个子集,对每个子集随机选择一个路段作为该子集所有路段的代表,从而构造路网数据的压缩矩阵。通过路网原始数据矩阵与构造的压缩矩阵计算两者之间的映射关系,对构造压缩矩阵的路段分别建立预测模型进行交通流预测,最后通过映射关系获得整个路网所有路段交通流的预测结果。实验结果表明提出的方法在保证预测精度的条件下,大大提高了运算速度,是一种有效的大规模路网交通流预测方法。更多还原

【Abstract】 With the development of social economy,the existing highway infrastructure has been difficult to meet people’s travel requirements.Traffic congestion on highways and urban roads has become a difficult problem for traffic management departments.With the rapid development of information,control and network technology,Intelligent Transportation Systems(ITS)are recognized as the main way to solve traffic congestion.In recent years,the emergence of new technologies such as Internet of things,Internet plus,big data and artificial intelligence has brought new vitality into ITS.It has become a hot topic in ITS research how to make use of the latest artificial intelligence technology to predict the traffic conditions,so that travelers can choose the right route,and reduce travel time,traffic congestion and environmental pollution.Based on the analysis of the research situation of traffic flow prediction at home and abroad,the spatiotemporal correlation characteristics of traffic flow are analyzed qualitatively and quantitatively with Pearson correlation coefficient and correlation function.Aiming at the time characteristic of traffic flow,the similarity and periodicity of traffic flow in working and non-working as well as the same continuous working days are analyzed,which provides a theoretical foundation for data set selection of traffic flow forecasting model.In terms of spatial characteristics,the correlation and local similarity of traffic flow in different sections are analyzed.The results show that the correlation of traffic flow in different sections is related to the distance between sections,but the correlation degree is not inversely proportional to the distance between sections,which provides a theoretical foundation for the construction of traffic flow prediction model.By analyzing the causes of abnormal traffic data,an abnormal traffic data detection method is proposed based on the local Leida criterion on the basis of analyzing the methods of Chauvenet,Dixon and Leida,and the experiments verify the reliability and effectiveness of the proposed method in abnormal traffic data detection.The reasons are analyzed for the lack of traffic data.According to determinate missing,a traffic data imputation method is proposed based on improved low rank matrix decomposition.The performance of the proposed method is tested and analyzed by the actual traffic flow data,which shows that the proposed method is effective.Based on the basic theory of discrete graph signal processing and the spatiotemporal correlation characteristics of traffic flow,a road network traffic flow adjacency matrix is constructed based on traffic data and distance correlation.A traffic flow decomposition algorithm is proposed for time series and road network based on Graph Fourier Transform,which decomposes traffic flow data into basic trend items and random fluctuations.Considering the temporal and spatial correlation of traffic flow,KNN algorithm is used to select related traffic flow data in different sections of the road network.Based on the decomposition of traffic flow data with Graph Fourier Transform,and the advantages of LSTM in processing time series,a hybrid traffic flow forecasting model called KNN-LSTM is proposed,which combines KNN,GFT and LSTM.The test results show that the proposed method has better prediction performance compared with classical ARIMA,SVR,ANN,DBN and LSTM models.In view of the current situation that there are few studies on traffic flow prediction in large-scale road network,a network traffic flow prediction method is proposed based on K-means clustering analysis and matrix decomposition.The traffic flow in different road sections is analyzed with K-means clustering in the road network,and the road sections are divided into K subsets using Euclidean distance as criterion.For each subset,a road section is randomly selected as the representative to construct the compression matrix of the road network data.The mapping relationship is calculated between the original road network data with the constructed compression matrix,and a traffic flow prediction model is constructed for selected sections.Finally,the predicted traffic flow in all sections is obtained through the mapping relationship in the whole road network.The experimental results show that the proposed method not only greatly improves the operation time,but also guarantees the prediction accuracy,and it is an effective method for large-scale road network traffic flow prediction.更多还原