随着中国城市化进程不断推进，交通拥堵等问题变得越来越严重，智能交通系统逐渐被应用于解决交通系统中的各类问题。交通流预测是通过道路特征、历史交通流量来预测未来的交通流。考虑到交通系统的复杂性，长时间的交通流预测很难达到较高的精度，因此，本文着力研究基于深度学习的短时交通流预测。本文系统全面地总结了深度学习模型在短时交通流预测领域的运用，并结合图卷积网络与长短时记忆网络，搭建了能够深层次提取交通流时空特征的混合模型LSTGCN，实现了精准的短时交通流预测。本文的研究内容主要从以下三方面工作展开：（1）基于路网交通流的转移效应与反馈效应，采用三种图卷积网络方法Chebnet、GCN、GAT对整个路网交通流的空间特征进行了提取，并基于上述方法搭建了交通流的图卷积模型，以道路邻接矩阵与历史交通流量作为输入，经过模型的训练学习得到了预测输出，对输出结果与真实数据进行了对比与误差分析，并分析了输入时间步、隐藏层节点数等因素对模型预测结果的影响。（2）基于路网交通流的周期性与趋势性，分别采用LSTM、GRU、Transformer对交通流的时间特征进行了提取，并基于三种网络分别搭建了交通流的时间预测模型，这里的输入仅为历史交通流量，但是取得的预测效果明显优于上面仅仅提取空间特征的深度学习模型，这说明了交通流数据具有显著的时域特征；同时分析了深度学习网络中的隐藏层节点数，历史交通流量输入时间步对输出结果的影响。（3）结合上述图卷积网络中的GCN模型与时间特征提取网络中的LSTM模型，搭建了能够同时提取交通流时间特征与空间特征的交通流时空混合预测模型长短时图卷积模型（LSTGCN），研究了输入时间步与LSTM隐藏层节点数对LSTGCN模型性能的影响，在深圳罗湖数据集与多种浅层模型、深层模型以及近几年来学者提出的T-GCN模型与STGCN模型进行了结果对比，预测误差有着显著的减少；最后通过两种方法验证了LSTGCN模型具有较强的鲁棒性，一是通过对深圳罗湖原始数据加入高斯噪声进行模型训练，发现误差变化极小，二是通过在洛杉矶数据集上进行LSTGCN模型的训练，得到了较好的预测结果。

With the continuous advancement of China's urbanization, traffic congestion and other problems become more and more serious. Intelligent transportation system is gradually applied to solve all kinds of problems in the transportation system. Traffic flow prediction is to predict the future traffic flow through road characteristics and historical traffic flow. Considering the complexity of traffic system, it is difficult to achieve high accuracy for long-term traffic flow prediction. Therefore, this paper focuses on short-term traffic flow prediction based on deep learning method.This paper systematically and comprehensively summarizes the application of deep learning method in the field of traffic flow prediction. Combined with graph convolution network and long-term and short-term memory network, a hybrid model LSTGCN which can deeply extract the temporal and spatial characteristics of traffic flow is built to realize accurate short-term traffic flow prediction. This paper mainly consists of the following three aspects:(1) Based on the transfer effect and feedback effect of road network traffic flow, three kinds of graph convolution networks Chebyshev Network(Chebnet), Graph Convolutional Network(GCN) and Grapg Attention Network(GAT) are used to extract the spatial characteristics of traffic flow respectively, and the graph convolution model of traffic flow is built based on the three networks. Taking the road adjacency matrix and historical traffic flow as the input, the prediction output is obtained through the training and learning of the model, and the output results are compared with the real data and the error is analyzed. The effects of input time steps and the number of hidden layer nodes on the prediction results of the model are analyzed.(2) Based on the periodicity and trend of road network traffic flow, the temporal characteristics of traffic flow are extracted by Long Short Term Memory(LSTM), Gated Recurrent Unit(GRU) and Transformer respectively, and the temporal prediction models of traffic flow are built based on the three networks. The input here is only historical traffic flow, but the prediction effect is obviously better than the deep learning model which only extracts spatial characteristics, which shows that the traffic flow data has significant time-domain characteristics; At the same time, the influence of the number of hidden layer nodes in the deep learning network and the input time step of historical traffic flow on the output results is analyzed.(3) Combining the GCN model in the above graph convolution network and the LSTM model in the temporal feature extraction network, a traffic flow spatiotemporal mixed prediction model long short term graph convolution network (LSTGCN) which can extract the temporal and spatial features of traffic flow at the same time is built. The effects of input time steps and the number of LSTM hidden layer nodes on the performance of lstgcn model are studied. The results of this model are compared with various shallow models, deep models, T-GCN model and STGCN model proposed by scholars in recent years on Shenzhen Luohu data set, and the prediction error is significantly reduced; finally, two methods are used to verify robustness of LSTGCN model. One is to add Gaussian perturbation to the original data of Shenzhen Luohu and find that the error change is very small, and the other is to train the LSTGCN model on the Los Angeles data set and prediction error is very small.