面向路网交通流态势预测的图神经网络模型

Graph Neural Network for Traffic Flow Situation Prediction

融合了路网结构的交通流态势预测是一个高度非线性化且复杂的时空动态相关性的时序数据预测问题。然而,传统交通流态势预测方法无法建模交通网络中长时间序列数据间的时空相关性。针对交通路网交通流态势预测问题,提出了一种基于图结构的交通流预测深度学习模型。首先,基于图小波变换定义图小波卷积算子,设计了面向路网交通流态势预测的图小波卷积神经网络模块;其次,结合时空注意机制构建了用于道路网络交通流态势预测的时空动态相关性模型,以捕获交通网络的动态时空相关性;最后,采用叠加多层图小波神经网络模块的策略,构建了一种面向路网交通流态势预测的图小波卷积神经网络模型。实验结果表明,该网络模型在数据集上的性能优于现有的基线模型。通过图小波变换矩阵与傅里叶变换矩阵非零元素统计对比实验,发现基于图小波变换定义的卷积运算更具稀疏性。因此,基于图小波变换定义的卷积运算更有助于提升交通流态势预测模型的计算效率。

Road network structure integrated traffic flow situation prediction is a highly nonlinear and complexly spatial-temporal dynamic correlation time-series data prediction problem.However,traditional traffic flow situation forecasting methods cannot model the temporal and spatial correlation in long-term series data in the traffic network.To address the issue mentioned above,a deep learning model of traffic flow prediction based on graph structure is proposed.Firstly,the graph wavelet convolution operator is defined based on the graph wavelet transform.Furthermore,the graph wavelet convolution neural network module is designed based on the operator for the traffic flow situation prediction.Secondly,a spatial-temporal dynamic correlation model is constructed based on the spatial-temporal attention mechanism to capture the dynamic temporal and spatial correlation of the traffic network.Finally,the strategy of stacking multi-layer graph wavelet neural network modules is adopted to establish a novel graph wavelet neural network for road network traffic flow situation prediction.Experimental results show that the developed model’s performance on the experimental datasets is better than the existing baseline models.In the comparative experiment on the non-zero element statistics of the graph wavelet transform matrix and the Fourier transform matrix,it's found that the convolution operation based on the graph wavelet transform is more sparse.Therefore,the convolution operation defined based on the graph wavelet transform is more helpful to improve the calculation efficiency of the traffic flow situation prediction model.