基于相互学习的短时交通流预测研究

Research on Short-term Traffic Flow Prediction Based on Mutual Learning

摘要:

交通流预测是智能交通系统（ITS）的核心，其中时空特性是最主要的特征。由于不同道路之间存在复杂的空间相关性和时间依赖性，因此交通流预测成为一项具有挑战性的任务。目前，基于图卷积神经网络的预测方法在网络局部以及整体的特征感知和提取方面，仍存在优化空间。为了解决以上问题，本文提出了一种基于图神经网络的优化模型：Diffusion Mutual Convolutional Recurrent Neural Network (DMCRNN)。该模型以DCRNN为基准模型，利用相互学习策略对其进行优化。在训练过程中，两个DCRNN网络之间相互学习、相互指导，以此来增强每个网络的特征学习能力。在METR-LA和PEMS-BAY两个真实数据集上验证优化策略的有效性。结果表明，经过优化后的模型预测误差显著降低，在两个数据集上一小时的MAE分别降低了0.15和0.12，即相互学习优化策略具有较好的性能。

Abstract:

Traffic flow prediction is the core of Intelligent Transportation Systems (ITS), with spatiotemporal characteristics being the most important feature. Due to the complex spatial correlations and time dependencies between different roads, traffic flow prediction has become a challenging task. Currently, prediction methods based on graph convolutional neural networks still have room for optimization in terms of feature perception and extraction at both local and global levels of the network. To address abpve issues, this paper proposes an optimized model based on graph neural networks: the Diffusion Mutual Convolutional Recurrent Neural Network (DMCRNN). This model is based on the DCRNN benchmark model and utilizes a mutual learning strategy to optimize it. During training, two DCRNN networks mutually learn from and guide each other to enhance their respective feature learning capabilities. The effectiveness of the optimization strategy is verified on two real datasets, METR-LA and PEMS-BAY. The results show that the optimized model significantly reduces prediction errors, with a decrease in MAE of 0.15 and 0.12 for one hour on the two datasets respectively, indicating that the mutual learning optimization strategy has good performance.