基于非线性交通流的深度时空图预测模型研究

Research on Deep Spatial-Temporal Graph Prediction Model Based on Non-Linear Traffic Flow

摘要: 在城市交通运营和管理中，快速、准确的交通流预测至关重要。传统的预测方法往往在处理由交通流的高非线性和复杂性带来的预测挑战方面存在不足，特别是在捕获交通流的动态空间关联性方面的缺失。图卷积网络被视为建立交通流的时空预测模型的关键技术之一。然而，在深度交通流时空预测模型中，这些模型往往面临过度平滑和梯度消失等问题。因此，本研究提出了一种深度时空图预测模型，重点解决了图卷积网络在交通流预测模型中的深度构建问题。该模型由一个结合了深度图卷积网络和时间卷积的时空预测模型组成，通过构建一种新式的可微广义聚合函数和残差连接策略，本研究实现了多层图卷积网络的信息聚合，并建立了一个能够捕获时间相关性特征的时间卷积模块。通过在真实交通数据集上的一系列实证测试，本研究验证了该模型的有效性超过了最新的基准方法，总体平均预测误差降低了30.4%。

Abstract: In urban traffic operation and management, swift and accurate traffic flow prediction holds immense importance. Traditional prediction methods often fall short in dealing with the prediction challenges brought by the high non-linearity and complexity of the traffic flow, especially lacking in deciphering the dynamic spatial correlations of the traffic flow. Graph convolutional networks are seen as the key technology to establish spatial-temporal prediction models for traffic flow. However, in deep traffic flow spatial-temporal prediction models, these models often face issues such as over-smoothing and gradient vanishing. Therefore, this study proposes a deep graph spatial-temporal prediction model, focusing on addressing the issue of the depth construction of graph convolutional networks in traffic flow prediction models. This model consists of a spatial-temporal prediction model that combines deep graph convolutional networks and temporal convolution. By constructing a differentiable generalized aggregation function and a residual connection strategy, this study realizes the information integration of multiple layers of the graph convolutional network. Simultaneously, by introducing the convolution mechanism, it establishes a temporal convolution module capable of capturing time-related characteristics. Through a series of empirical tests on genuine traffic datasets, this study validates that the model effectiveness exceeds the latest benchmark methods, and on the whole, the predictive error is reduced within the range of 30.4%.