网约车需求时空预测——以神州专车为例

论文总字数：34106字

摘 要

随移动互联网的普及和发展，许多行业都在“互联网 ”的影响下实现了升级发展。传统的巡游出租车长期以来都由于司机和乘客的信息不对称导致司机乘客难以匹配、乘客需求难以被满足，因而巡游出租车行业也理所当然地受到移动互联网的影响。网约车即是“互联网 传统巡游出租车”而产生的。网约车的司机和乘客双方在网约车平台上实现双向搜寻，乘客可以在网约车平台上提出需求，而司机则能在平台上搜寻乘客，这使得乘客需求能更快的被司机接收并得到满足，司机与乘客匹配的效率大大提高。虽然网约车提高了司机和用户的匹配效率，但在司机和乘客的时空分布极度不平衡时，乘客叫不到车和司机找不到乘客的情况还是很有可能发现。为了协调尽可能地提高司机和乘客的匹配成功率，需要提前预测乘客需求的时空分布，然后通过优惠奖励等方法将空闲车辆提前调配的乘客需求量高的区域，减少乘客打不到车和司机空车的情况，提高资源利用效率和用户满意度。

本文根据以往研究将历史需求量、天气特征变量、时间特征变量、POI数量、平均距离价格比以及平均距离时间比作为网约车时空需求的可能影响因素并用信息增益的方法进行特征变量筛选。然后分别用基于最小二乘的多元线性回归模型和基于长短期记忆网络的深度学习模型对乘客需求进行预测，并对预测结果进行分析评价和比较。

关键词：网约车，需求量预测，多元线性回归，LSTM

Abstract

For a long time, the information asymmetry between drivers and passengers has made it difficult for drivers and passengers to match and passengers' needs to be satisfied. Therefore, with the popularity and development of mobile Internet, the cruise taxi industry has also been affected. Ride-hailing is the result of "Internet traditional cruise taxi".Both drivers and passengers of ride-hailing services can realize two-way search on the platform. Passengers can put forward demands on the platform while drivers can search passengers on the platform. This makes the demands of passengers more quickly satisfied by drivers and greatly improves the matching efficiency between drivers and passengers. Although ride-hailing improves the matching efficiency between drivers and passengers, it is likely to be found when passengers cannot hail a car and drivers cannot find passengers when the spatial and temporal distribution between drivers and passengers is extremely unbalanced. To coordinate as much as possible to improve the matching success rate of drivers and passengers, the need to forecast passenger demand of time and space distribution in advance, and then through such means as preferential rewards will spare vehicles deploy in advance to the passenger areas of high demand, reduce passengers could not hit the car and driver empty, improve resource utilization efficiency and user satisfaction.

Based on previous studies, this paper takes historical demand, weather characteristic variables, time characteristic variables, POI quantity as possible influencing factors for the space-time demand of ride-hailing vehicles and selects characteristic variables by information gain method. Then, the passenger demand is predicted by the multiple linear regression model based on least square and the deep learning model based on long and short term memory network respectively, and the predicted results are analyzed, evaluated and compared.

KEY WORDS: ride-hailing, demand forecasting, multivariate linear regression, LSTM

目 录

摘 要 3

Abstract 4

第一章 绪论 1

1.1 选题背景和意义 1

1.2 本文组织结构 2

第二章 文献综述 3

第三章 研究方法 5

3.1 LSTM神经网络模型 5

3.1.1 循环神经网络 5

3.1.2 长短期记忆网络（long short-term memory，LSTM） 6

3.2 多元线性回归模型 7

3.3 特征选择方法——信息增益 7

第四章 数据源说明与处理 9

4.1 数据基本特征说明 9

4.2 数据预处理 10

4.2.1 时间段划分与天气数据选择 10

4.2.2 区域编码 11

4.2.3 重复数据与缺失数据处理 11

第五章 变量特征分析 12

5.1 时间特征与需求量的关系 12

5.2 POI数量与需求量的关系 15

第六章 网约车需求时空预测 20

6.1 变量定义解释 20

6.2 特征选择 21

6.2.1 筛选出相关性过大的特征 21

6.2.2 根据信息增益筛选变量 22

6.2.3 检查变量间的多重共线性 23

6.2.4 通过多元线性回归进一步筛选变量 23

6.3 实验数据的选取与训练集、测试集的划分 25

6.4 多元线性回归模型求解与预测 25

6.4.1 多元线性回归模型求解 25

6.4.2 多元线性回归模型预测 26

6.5 神经网络模型预测 27

6.5.1 模型描述 27

6.5.2 模型训练与预测 27

6.6 模型比较 27

第七章 总结与展望 28

7.1 全文总结 28

7.2 存在的问题与改进方向 28

参考文献 29

附录 32

剩余内容已隐藏，请支付后下载全文，论文总字数：34106字