基于深度学习的短时交通流预测

发布时间：2018-03-14 04:19

本文选题：短时交通流预测　切入点：深度学习　出处：《青岛大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着城市智能交通系统的不断完善,城市道路上安装了大量的检测器设备来采集道路和车辆的信息数据,通过分析检测器检测到的大量数据,为从数据中发现交通规律,支持城市交通优化,短时交通流分析是当前主要研究问题之一。卷积神经网络避免了从数据中人工提取特征的问题,采用卷积神经网络科自动提取交通流量的时空特征,具有较好的应用价值。基于此本文采用卷积神经网络的方法对短时交通流进行研究预测,主要内容如下:本文通过构建卷积神经网络模型,将路网的交通流数据转化为交通流拥堵级别数据,实现基于此类数据的预测时间小于5分钟的交通状况预测,面对特定类型的样本较小的情况,利用迁移学习的思路增加训练集的数据量和提升模型的预测性能。本文的主要工作如下:(1)针对短时交通流预测问题,提出了采用卷积神经网络解决,分析了方案中卷积神经网络中卷积核大小对模型预测准确度的影响,采用秦皇岛连续并且缺失值最少的15天的数据集,实现了1分钟和5分钟的流量预测。并实现了高峰期和全天时段数据的预测。(2)针对数据训练样本过少,导致模型过拟合问题,采用迁移学习的思想,通过不同时段数据集的不同以及在训练集中增加随机扰动的方法,增大了数据样本,提高了模型预测准确性。
[Abstract]:With the continuous improvement of urban intelligent transportation system, a large number of detectors are installed on urban roads to collect information data of roads and vehicles. In order to support urban traffic optimization, short-term traffic flow analysis is one of the main research problems. Convolution neural network avoids the problem of manually extracting features from the data, and uses convolution neural network to automatically extract the space-time characteristics of traffic flow. Based on this, the method of convolution neural network is used to study and predict the short-term traffic flow. The main contents are as follows: this paper constructs the model of convolution neural network. The traffic flow data of the road network are transformed into traffic congestion level data, and the traffic condition prediction with less than 5 minutes prediction time based on this kind of data is realized. In the case of small sample size of specific type, the traffic flow data of the road network is transformed into the traffic congestion level data. The main work of this paper is as follows: (1) to solve the problem of short-term traffic flow prediction, a convolution neural network is proposed to solve the problem. The effect of convolution kernel size on prediction accuracy in convolution neural network is analyzed. The data set of Qinhuangdao continuous with the least missing value for 15 days is used. 1 minute and 5 minute traffic forecasting is realized. The prediction of peak period and whole day data is realized. In view of too few data training samples, the model is overfitted and the idea of transfer learning is adopted. Through the different data sets in different periods and the method of adding random disturbance in the training set, the data samples are enlarged and the prediction accuracy of the model is improved.
【学位授予单位】：青岛大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.14;TP18

【参考文献】