基于离散小波变换的深度树时空模型

A Deep Spatiotemporal Model for Traffic Flow Prediction Based on Discrete Wavelet Transform

摘要: 交通流量预测是智能交通领域的研究热点，有利于提高交通资源分配的合理性和出行政策制定的有效性。新型冠状病毒的爆发，严重影响了交通出行的正常秩序。许多国家为了减少疫情的传播速度均颁布了限制居民出行的政策，导致交通流量数据出现了较高的离散性和不规则性。为了克服数据离散性对交通流量预测的影响，本文采用离散小波变换(DWT)将交通流量数据拆分为离散量、变化趋势和离散基线。为了提升模型对高离散性交通数据预测的准确率，本文设计了两种不同的模型来分别预测变化趋势和离散量。由于出行区域的限定，疫情期间的交通状况变化呈现出小规模聚集性。图卷积神经网络的节点邻接计算方法适用于节点随机均匀分布的图结构，对于节点小规模聚集分布的图结构的计算效果较差。本文提出了一种树卷积网络(TreeCN)来分析交通网络的空间相关性，并采用时序卷积网络来分析交通数据的时间相关性。为了解决交通流量数据的高离散性问题，本文提出了一个离散预测模块(DPM)，用于将离散小波变换分离出的离散量转换为高维离散特征。最后，使用离散小波变换对预测的交通数据进行分割，然后将新分割的交通趋势和离散基线与离散预测模块预测的离散模型进行逆离散小波变换，得到最终的交通流预测结果。在对比实验中，将这项工作与现有的高级基线进行了比较，本文模型要优于现有基线模型。

Abstract: Traffic flow prediction is a focal point in the field of intelligent transportation, aiming to enhance the rationality of traffic resource allocation and the effectiveness of travel policy formulation. The out-break of the novel coronavirus has significantly disrupted the normal order of traffic movements. To mitigate the impact of data discreteness on traffic flow prediction caused by policies restricting resident mobility during the pandemic, this study employs Discrete Wavelet Transform (DWT) to decompose traffic flow data into discrete components, trend variations, and discrete baselines. To improve the accuracy of predicting high discrepancy traffic data, two distinct models are designed to forecast trend variations and discrete components separately. Due to restricted travel areas, the traffic conditions during the pandemic exhibit small-scale clustering. The conventional node adja-cency calculation method of Graph Convolutional Neural Networks (GCNN) is suitable for graph structures with nodes randomly and uniformly distributed, but it performs poorly for graph struc-tures with nodes exhibiting small-scale clustering. This paper proposes a Tree Convolutional Net-work (TreeCN) to analyze the spatial correlations of the traffic network and utilizes a Temporal Convolutional Network to analyze the temporal correlations of traffic data. To address the high dis-creteness issue in traffic flow data, a Discrete Prediction Module (DPM) is introduced to transform the discrete components extracted by the discrete wavelet transform into high-dimensional dis-crete features. Finally, the study utilizes discrete wavelet transform to segment the predicted traffic data, and then combines the newly segmented traffic trends and discrete baselines with the dis-crete model predicted by the Discrete Prediction Module through inverse discrete wavelet trans-form, yielding the ultimate traffic flow prediction results. Comparative experiments demonstrate that the proposed model outperforms existing advanced baseline models.