毕业生知识库

交通是城市的纽带，在人们的生产生活中扮演着举足轻重的角色。然而，随着城市化进程的加快，城市交通所面临的压力越来越大，各种问题日渐凸显。交通拥堵、交通资源供需不平衡和交通事故等问题严重影响了人们的出行质量。从人们的出行数据中挖掘规律，并由此制定合理的对策，可有望缓解以上问题。
城市出行起止需求预测通过研究城市区域间不同时段的需求数量来揭示人们的出行规律。准确的出行起止需求预测对于交通管理具有重要意义。技术上，如何挖掘隐藏在出行起止需求数据中的复杂时空关联关系是模型设计的重点。传统的统计模型和基于循环神经网络的模型对单个节点的历史序列建模，从中挖掘出行起止需求数据中的时间相关性。但是，这种方式忽略了节点间的空间相关性。基于卷积神经网络的模型将城市划分为栅格，将起止需求数据构造成多通道的栅格数据，然后使用卷积神经网络建模区域间的空间相关性。但是，这种方法难以直接应用于地铁和公交出行数据。近年来，基于图卷积神经网络的预测方法广泛应用于交通预测任务中。这种方法首先将交通网络建模为节点和边组成的图，然后定义节点的特征，接着使用图卷积网络来建模节点间的空间相关性。但是，这种方法主要用于捕捉相邻节点之间的相关性，难以建模距离较远的节点之间的联系。城市出行起止需求预测需要分析城市内远距离区域之间的关联关系。相对来讲，相关研究工作仍不丰富，在模型精度、推理速度等方面尚不能满足应用需求。为此，本文开展基于图卷积神经网络的城市出行起止需求预测方法研究和系统研制工作。本文的主要工作如下： 
    1.提出了一种基于图神经网络的出行起止需求预测模型（Graph-attention LSTM Network, GLNet）。模型的整体架构采用了序列预测任务中常用的编解码框架。在空间相关性建模方面，GLNet使用图注意力机制来建模节点间的空间相关性。针对图注意力网络缺乏长距离建模能力的问题，本工作设计了一个自适应的邻接矩阵来建模节点间的连接关系，并将其参与到注意力分数的计算中。在时间相关性建模方面，GLNet使用长短时记忆网络（Long Short Term Memory，LSTM）作为编码器来建模出行起止需求数据的时间相关性。针对编码器初始化状态没有利用到先验知识的问题，本工作设计了一个隐状态初始化模块，用于学习编码器的初始状态。另外，GLNet额外引入了一个历史编码损失（Historical Encoding Loss，HELoss），通过降低历史隐状态的解码结果与真实值的差距，提升了模型的预测能力。最后，本工作构造了三个专有数据集，并在此数据集上进行了实验。GLNet在MAE、RMSE等多个指标上表现优异，验证了所提方法的有效性。
    2.开发了一个基于浏览器/服务器（Browser/Server，B/S）架构的出行起止需求可视化系统，并将其用于呈现和分析北京公交1路的出行情况。具体地，出行起止需求可视化系统集成了数据展示模块和预测模块。数据展示模块使用多种可视化图表从不同角度对出行起止需求数据进行展示和分析，帮助人们发现交通出行规律。预测模块则集成了包含本文所提方法在内的多种预测模型，用来预测未来半小时内的出行起止需求。

Traffic acts as fundamental ties for cities, and plays an important role in goods transportation and passenger travel. However, the pressures on urban transportation are increasing with the acceleration of urbanization, carrying out various types of problems to be solved. Traffic congestion, imbalanced supply and demand of transportation resources, and traffic accidents have seriously affected travel quality. Mining patterns from travel data to make reasonable plans and decisions can effectively alleviate the aforementioned problems.

By studying the number of demands between urban regions at different times, urban travel origin-destination demand forecasting can reveal travel patterns. Accurate travel demand forecasting is of great significance for traffic management. Technically, how to explore the complex spatiotemporal correlations hidden in travel origin-destination demand data is the key of model design. Traditional statistical models and models based on Recurrent Neural Networks(RNN) model the historical sequence of a single node, mining the time correlation in travel demand data. However, this approach ignores the spatial correlation between nodes. Models based on Convolutional Neural Networks(CNN) divides city into grids, constructs origin-destination demand data into multi-channel grid data, and then uses CNN to model spatial correlations between grids. However, these methods are difficult to be directly applied to subway and bus travel data which are organized with graphs. In recent years, Graph Convolutional Neural Networks (GCNN) have been widely used in traffic prediction. This method first models the traffic network as a graph composed of nodes and edges, then defines the characteristics of the nodes, and finally uses a GCNN to model the spatial correlation between nodes. However, existing GCNN methods is mainly used to capture the correlation between adjacent nodes, and have limited capabilities to model the connection between distant nodes. Technically, urban travel origin-destination demand forecasting requires analyzing the correlation between distant regions within a city. Relatively speaking, the related research works in the literature are not rich, and accuracy and inference speed obtained by existing models cannot meet the requirements in real-world applications. To this end, this thesis carries out the research and system development for urban travel demand forecasting with GCNN tricks. The main work of this thesis is as follows:
    1. A travel origin-destination demand forecasting model with graph neural networks (Graph-attention LSTM Network, GLNet) is proposed. The overall architecture of the model adopts the commonly used encoding and decoding frameworks in sequence prediction tasks. For spatial correlation modeling, GLNet uses graph attention mechanism to model the spatial correlation between nodes. To address the disadvantage of graph attention networks that lack the capabilities of long-distance modeling, this thesis designs an adaptive adjacency matrix to capture the spatial relationships between nodes and uses it for attention score estimation. For temporal correlation modeling, GLNet uses Long Short Term Memory(LSTM) Network as the encoder to model the time correlation of travel demand data. To address the issue of the encoder initialization state not utilizing prior knowledge, this thesis designs a hidden state initialization module for learning the initial state of the encoder. Furthermore, GLNet introduces an additional Historical Encoding Loss (HELoss), which improves the predictive ability of the model by reducing the difference between the decoding results of historical hidden states and the true values. Finally, this thesis constructed three proprietary datasets and conducted experiments on them. GLNet performs well on multiple indicators such as MAE and RMSE, verifying the effectiveness of the proposed method.
    2. A travel origin-destination demand visualization system is constructed with Browser/Server (B/S) architecture, which is employed to present and analyze the travel situation of Beijing Bus Route 1. Specifically, the travel origin-destination demand visualization system integrates a data display module and a prediction module. The data display module uses various visualization charts to display and analyze travel origin-destination demand data from different perspectives, which helps to discover traffic travel patterns. The prediction module integrates multiple prediction models to predict travel origin-destination demand for the next half hour.