【摘要】 针对现有短时交通预测模型主要关注交通速度，忽视时段特性、天气和道路结构等因素的问题，提出了基于多源数据的动态时空图卷积神经网络模型（MDSCNN）。该模型全面整合多源特征矩阵、邻接矩阵和相似矩阵，通过梯度下降法构建可更新混合矩阵，结合图卷积、长短期记忆网络和注意力机制，进行多时段短时交通速度预测。多个数据集的实验结果表明，MDSCNN在各方面均优于其它模型，并通过外源因素对比实验验证了多源数据的引入能够有效提升预测准确性。更多还原

【Abstract】 To overcome the limitations of existing short-term traffic prediction models that mainly focus on traffic speed while neglecting temporal patterns, weather conditions, and road structure, a dynamic spatiotemporal graph convolutional neural network(MDSCNN) based on multi-source data was proposed. Multi-source feature matrices, adjacency matrices, and similarity matrices were comprehensively integrated by this model. An updatable hybrid matrix was constructed via gradient descent, combining graph convolution, long short-term memory networks, and attention mechanisms for multi-period short-term traffic speed prediction. Experimental results on multiple datasets demonstrate that MDSCNN outperforms other models in all aspects, and comparative experiments on exogenous factors verify that the introduction of multi-source data effectively improves prediction accuracy.更多还原