基于轨迹数据挖掘的短时出租车区域分布预测研究

发布时间：2018-05-16 00:19

本文选题：出租车 + 轨迹数据挖掘　；参考：《吉林大学》2017年硕士论文

【摘要】：出租车是城市公共交通系统中一个不可或缺的重要组成部分,为市民提供了一种方便快捷的出行方式。近年来,随着移动互联网的发展,类似于滴滴这样的网约车服务平台兴起,一方面在为市民提供了一种更加高效便捷的出行方式选择的基础上,另一方面也对传统的出租车运营模式造成了一定程度的冲击,传统的出租车行业面临较大的竞争压力。出租车运营公司迫切需要提升运营效率提高竞争力,抗衡网约车所造成的冲击。其中有效的空载出租车运力调度是降低出租车空载率,提升运营效率的关键,而科学合理的空载出租车调度基于对未来一段时间内不同区域打车需求情况以及出租车运营公司旗下所有出租车在不同区域的分布情况的精确预测,打车需求远大于出租车供应的区域是出租车调度的目的地,通过将未来会出现在供大于求区域的空载出租车调度到未来出租车供小于求的区域,从而达到出租车供需平衡的状态是空载出租车调度所追求的目标。本文主要关注于对未来一段时间出租车的区域分布情况进行预测的研究。与打车请求所具有的明显的随机性不同,未来一段时间内出租车在不同区域的分布情况与当前出租车在各个区域的分布情况高度相关,因此对出租车区域分布进行预测主要利用当前出租车的区域分布信息。通过挖掘历史出租车轨迹数据,本文提出了多个不同类型的短时出租车区域分布预测算法,并且通过模拟预测实验,验证对比了各个算法的预测效果,三种预测算法分别是基于概率统计的马尔可夫过程预测算法,属于无监督学习的矩阵分解算法,属于监督学习的GBRT预测算法等。马尔可夫过程是随机过程的一种,其最重要的性质是马尔可夫性即无后效性,在短时出租车区域分布预测问题中,基于马尔可夫过程的预测算法通过将实时出租车区域分布抽象为向量形式,然后与描述一天中此时段出租车在各个区域之间进行转移的区域转移概率矩阵进行矩阵乘法运算,从而获取一段时间后出租车在各个区域内的分布预测。矩阵分解是隐语义模型算法的核心步骤,主要应用于推荐系统领域,本文中将矩阵分解算法引入出租车区域分布预测问题中,并且基于出租车区域分布的时空特性对基本的矩阵分解算法应用进行了改造,使其可以有效的适用于出租车区域分布预测问题。GBRT算法是一种典型的监督机器学习算法,具有泛化能力强,预测精确度高的特性,本文中通过为每一个区域单独训练一个回归器的方法来预测每个区域内一段时间后会出现的出租车数目。为了有效的利用出租车运营公司多年来所积攒的海量轨迹数据,本文中利用轨迹数据挖掘技术对原始的轨迹数据进行了预处理,从原始的轨迹数据中抽取出了与出租车时空分布相关的信息,并且将其组织为Tensor形式。在随后的预测算法学习与模拟预测阶段,可以方便的将Tensor中保存的相关信息转换为适合学习算法进行训练与模拟预测的形式。
[Abstract]:Taxi is an indispensable and important part of the urban public transportation system, which provides a convenient and quick way for the citizens to travel. In recent years, with the development of mobile Internet, the network of car service platform, which is similar to drip, has been rising. On the one hand, it provides a more efficient and convenient way for the citizens to choose the way to travel. On the other hand, on the other hand, the traditional taxi operation mode has caused a certain degree of impact, the traditional taxi industry is facing greater competition pressure. The taxi operation company urgently needs to improve the operation efficiency to improve the competitiveness and counterbalance the impact caused by the network about the car. Car rental rate is the key to improve the operation efficiency, and the scientific and reasonable taxi dispatch is based on the demand for the taxi in different areas in the next period of time and the accurate prediction of the distribution of all taxis under the taxi operation company in different areas. The taxi demand is far greater than the taxi supply area is the taxi regulation. In order to achieve a taxi supply and demand balance, the aim of the taxi dispatching is to dispatch the future taxi to the area where the taxi supply is less than the demand in the future. This paper mainly focuses on the prediction of the regional distribution of the taxi in the future period. The distribution of taxis in different regions is highly related to the distribution of taxis in different regions in the next period of time. Therefore, the regional distribution of taxis is predicted mainly by the regional distribution information of current taxis. Through mining historical taxis In this paper, a number of different types of short-term taxi regional distribution prediction algorithms are proposed, and the prediction results of each algorithm are compared by simulation prediction experiments. The three prediction algorithms are Markov process forecasting based on probability statistics, which belong to the unsupervised learning matrix decomposition algorithm and belong to the supervision. The GBRT prediction algorithm for learning. The Markov process is one of the random processes. The most important nature of the process is that the Markov nature is no aftereffect. In the short time taxi regional distribution prediction problem, the prediction algorithm based on the Markov process is abstracted into the vector form by the real-time taxi area distribution, and then it is described in the middle of the day. In order to obtain the distribution prediction of taxis in various regions after a period of time, a taxi can obtain the distribution prediction of the taxis in each region. Matrix decomposition is the core step of the algorithm of the semantic model of the hidden language, which is mainly used in the field of recommendation system. In this paper, the matrix decomposition algorithm is introduced to rent. In the vehicle regional distribution prediction problem, and based on the spatial and temporal characteristics of the taxis distribution, the application of the basic matrix decomposition algorithm is reformed, so that it can be effectively applied to the taxi regional distribution prediction problem.GBRT algorithm is a typical supervised machine learning algorithm, which has the characteristics of strong generalization ability and high prediction accuracy. In this paper, the number of taxis that will appear after a period of time in each region is predicted by training a regression device separately for each region. In order to effectively use the massive trajectory data accumulated by the taxi operators for many years, the trajectory data mining technique is used to preprocess the original trajectory data in this paper. The information related to the space-time distribution of taxis is extracted from the original trajectory data, and it is organized into a Tensor form. In the subsequent prediction algorithm learning and simulation prediction phase, the related information stored in the Tensor can be easily converted into the form of training and simulation prediction suitable for learning algorithms.

【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F570;O211.62

【参考文献】