基于分段平稳时间序列分析的高速公路交通流预测研究

Rencently,Traffic safety and congestion is a major challange in china, and traffic flow forecasting is an effective way to alleviate this problem. However, restricted by complexity and time-varying factors of traffic flow, the prediction accuracy needs to be improved urgently. Base on the research result of applicant before on short time traffic flow forecasting, the mid-term prediction of traffic flow would be explored in this project, which will provide an important technical support for real time traffic flow monitoring. Based on piecewise stationary time series analysis, then two algorithms will be presented, that is a novel traffic density prediction algorithm based on time series analysis and traffic flow rate forecasting algorithm based on piecewise homogeneous possion process.. In the aspect of traffic density prediction algorithm, on the basis of time series analysis and AR model, a novel online breaks estimation method will be proposed first.The parameters of AR model are calculated by AIC(Akaike information criterion) and the regression coefficient estimates method. Then, forecasting revisions will be processed by Kalman filter combined with AR model. Generally, in order to predict the traffic density, the non-stationary time series are divided into several piecewise stationary stochastic signal, and the forecasting effect will be improved.. In the aspect of traffic flow rate forecasting alorithm, on the basis of piecewise stationary time series anslyasis and piecewise homogeneous poisson process,assuming that different freeway sections are shown to be poisson process with different intensity parameters in different time, that is the traffic flow rate is piecewise stationary time series within certain time and length. Based on the real-time and history traffic data, the intensity parameters is estimated and the perturbation analysis is given. After several predictions training, the perturbation terms are optimized and its precision will be improved, and then the mid-term forecasting of freeway traffic flow rate wil be given. In addition, this algorithm could be used to fill the missing data of monitor points. . Finally, an experimental simulation platform base on domestic and international traffic flow data will be constructed, the algorithm above will be validated with three types of data, that is weekday data, weekend data and major holiday data..

交通安全与交通拥堵是目前我国面临的重大挑战，交通流预测是缓解该问题的有效途径之一。然而，交通流的复杂性、时变性等因素导致该项技术的准确性亟待提高。本项目在申请人从事短时交通流预测研究工作的基础上，对高速公路交通流中期预测进行探索，从而为基于实时路况的交通流监控提供重要技术支撑。该项目基于分段平稳时间序列分析的思想，假设高速公路交通流在某一时段内或某一空间区域内是平稳随机信号，继而构建基于时间序列分析的中期车流密度估计与预测算法，提出一种适合中期预测的在线式断点估计方法，并与多状态转移Kalman滤波相结合；构建基于分段时齐泊松过程的中期交通流量估计与预测算法，针对实时数据和历史数据，引入强度参数估计与摄动分析，并给出不同的影响因子，以提高预测精度。最后，本项目将构建试验性仿真验证平台，并以国内高速公路实测数据和美国加州PeMS平台共享交通流数据为基础，对上述预测机制与算法进行验证。

高速公路交通流预测在缓解拥堵、减少交通事故等方面起到非常重要的作用。囿于交通流的时变性和复杂性，预测精度亟待提高。由此，项目开展了一维参量的交通流预测和多维交通流参量估计工作。首先对高速公路车流密度和交通流分别进行预测研究；其次，为了将“能见度”引入交通流预测，基于变分和总有界变差(TBV)对能见度估计展开研究。概述如下：(1) 提出一种基于时间序列的车流密度预测方法。首先设计一种基于多参量的断点估计方法，得到分段平稳信号的断点位置和数量；继而估计AR模型参量并校正车流密度。(2) 提出一种基于分段时齐泊松过程的车流量预测方法。假定不同路段的交通流量在不同时段服从不同强度参数的泊松过程。通过多断面的交通流数据监测，估算出泊松过程的强度参数，并以下一断面的实测数据和摄动分析对预测值进行修正。本项目基于连续15年(2003-2017)的交通流实测数据(PeMS)对算法(1)和(2)进行验证，前者误差(MAPE)在7.9%至14.45%之间，后者在7.05%至14.29%之间。(3) 提出一种基于变分的能见度估计方法，为交通流预测提供新维度的参量。拓展消光系数为时间的函数，构建泛函并求其变分；利用分段平稳思想对亮度曲线进行逼近求消光系数，并构建分段校准函数，以估计能见度。(4) 提出一种基于TBV和图像频谱的低能见度(<300 m)估计方法。将雾霾视为清晰图像上叠加的噪声，以TBV和图像频谱提取图像特征。对于300米以下雾霾图片，其余弦变换高频系数比值小于20%，低频系数比值在100%-120%之间。利用这一特征搜索低能见度图片，继而构建TBV与实际能见度之间的函数关系。项目基于江苏省通启高速公路的实测视频数据(~200万帧)，对算法(3)(4)进行验证，误差均小于10%。