基于机器学习方法的城市对外客运交通需求预测研究

发布时间：2018-01-07 22:35

本文关键词：基于机器学习方法的城市对外客运交通需求预测研究　出处：《哈尔滨工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：城市对外客运交通需求预测是城市开展城市综合交通系统规划与设计的基础工作,合理准确的交通需求预测可为城市的对外客运枢纽系统选址、布局、方案比选等工作提供数据支撑,实现既满足城市居民出行需求,又节约项目建设资金的目标。由于对外客运需求预测研究中相关影响因素之间存在日趋增加的相关性关系以及统计数据中的异常值等原因,传统的时间惯性与相关因素原理预测模型表现欠佳。近几年由于社会统计工作的逐渐完善,可供选择研究统计数据不断积累增多,为学者使用新型方法进行研究提供了相关基础。本文采用机器学习中降噪自编码、随机森林两种方法进行交通需求预测,以缓解浅层机器学习方法在交通需求预测问题中的不足。首先引入深度学习理论中降噪自编码方法:降噪自编码方法通过数据的逐层自编码、解码过程获得良好的交通需求预测网络初始化参数,使得网络初始总体损失值较优,缓解了浅层需求预测方法的局部极值与梯度弥散问题。此外人工主动随机噪声,迫使网络在输入包含噪声的情况下重构原始输入,进而训练所得交通需求预测网络鲁棒性、泛化能力更强,不易过拟合。另外考虑对外客运出行需求的相关影响因素间的关联性和时间惯性,将时间序列数据研究中的窗口滑移与机器学习中的随机森林方法相结合,提出时间窗-随机森林组合方法的对外客运总体需求预测方法。随机森林方法在训练过程中共进行两重随机过程,第一重随机为在宏观交通相关数据总体训练样本中随机抽取部分样本训练决策树模型,未被抽取数据用以评价所得交通需求决策树预测模型泛化性能,多次随机抽样获得多颗决策树构成交通需求预测森林模型;第二重随机为在单棵决策树节点分裂过程中随机选取部分属性。两重随机过程使得模型过度拟合特定样本的概率大大减少,预测模型的泛化性增强。同时以北京市宏观经济影响因素数据集为基础进行实例分析,模型精度良好,验证了方法的可行性和有效性,可运用于对外客运需求预测工作。本研究侧重基于机器学习方法的对外客运需求预测,分别从方法由来、数学原理与方法实现等方面进行了详细阐述,可对省份、城市等范围区域进行交通运输发展规划研究工作提供参考与借鉴。对机器学习理论与交通问题的结合有着积极的作用。
[Abstract]:Urban external passenger transport demand prediction is the basic work of urban comprehensive transportation system planning and design. Reasonable and accurate traffic demand prediction can be used for the location and layout of urban external passenger transport hub system. The scheme provides data support to meet the travel needs of urban residents. The goal of saving project construction funds. Due to the increasing correlation between the related factors and the abnormal value in the statistical data in the forecast study of passenger demand for foreign passenger transport, and so on. The traditional prediction model of time inertia and related factors is not good. In recent years, due to the gradual improvement of social statistics, it is possible to choose to study the statistical data accumulation and increase. In this paper, two methods of noise reduction in machine learning and stochastic forest are used to forecast traffic demand. In order to alleviate the deficiency of shallow machine learning method in traffic demand prediction problem. Firstly, the noise reduction self-coding method is introduced in depth learning theory: noise reduction self-coding method through the data layer by layer self-coding. In the decoding process, good traffic demand prediction network initialization parameters are obtained, which makes the initial total loss value of the network better. The problem of local extremum and gradient dispersion of shallow demand prediction method is alleviated. In addition, artificial active random noise forces the network to reconstruct the original input when the input contains noise. Furthermore, the trained traffic demand forecasting network is robust, more generalized and difficult to be over-fitted. In addition, the correlation and time inertia among the related factors of external passenger travel demand are considered. The window slippage in time series data is combined with the stochastic forest method in machine learning. A time window-stochastic forest combination method is proposed to predict the total demand of passenger transport. The stochastic forest method carries out double stochastic processes during the training process. The first is random training decision tree model which is randomly selected from the total training samples of macro-traffic related data, and is not extracted to evaluate the generalization performance of the traffic demand decision tree prediction model. Multiple random sampling to obtain multiple decision trees to form a forest model of traffic demand prediction; The second random is the random selection of some attributes in the split process of a single decision tree node. The probability of overfitting a particular sample is greatly reduced by the double random process. The generalization of the prediction model is enhanced. At the same time, based on the data set of the macroeconomic impact factors in Beijing, the model has good accuracy, which verifies the feasibility and effectiveness of the method. This research focuses on forecasting the demand of foreign passenger transport based on machine learning method, respectively from the origin of the method, mathematical principles and the realization of the method are described in detail. It can be used as a reference for the study of transportation development planning in provinces, cities and other areas, and has a positive effect on the combination of machine learning theory and traffic problems.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U12

【参考文献】