The invention discloses a multi-head attention traffic flow prediction method based on a space-time diagram ordinary differential equation. The method comprises the following steps: S1, constructing three types of adjacent matrixes based on spatial correlation, semantic similarity and space-time correlation of traffic flow; s2, the three types of adjacent matrixes are input into STAODE in parallel, and three independent channels are formed in an STAODE layer; s3, information of three different blocks is generated through the STGODE modules of the three different channels; s4, aggregating different information of each channel through maximum pooling operation, and inputting the information into two layers of multi-layer perceptron layers (MLP); and S5, finally performing prediction through a multi-head attention mechanism. According to the method, the relation between time correlation and spatial correlation between data is considered, a new spatial adjacency matrix is added, a more comprehensive spatial relation is established, and meanwhile, a multi-head attention mechanism is introduced to strengthen the prediction effect. The method is obviously superior to other baseline models and has space-time effectiveness.
本发明公开了一种基于时空图常微分方程的多头注意力交通流预测方法,以下步骤:S1:分别基于交通流量的空间相关性、语义相似性和时空相关性构建三种类型的邻接矩阵;S2:将三种类型的邻接矩阵平行地输入到STAODE中,在STAODE层形成三个单独的通道;S3:通过三个不同通道的STGODE模块生成三个不同块的信息;S4:再通过最大池化操作聚合每个通道的不同信息,并输入到两层多层感知机层MLP中;S5:最后通过一个多头注意力机制进行预测。本方法考虑到数据之间的时间相关性和空间相关性之间的联系,加入新空间邻接矩阵,建立更加综合的空间关系,同时引入多头注意力机制强化预测效果。本方法明显要优于其他基线模型,更具有时空有效性。
Multi-head attention traffic flow prediction method based on time-space diagram ordinary differential equation
基于时空图常微分方程的多头注意力交通流预测方法
2024-05-24
Patent
Electronic Resource
Chinese
IPC: | G06Q Datenverarbeitungssysteme oder -verfahren, besonders angepasst an verwaltungstechnische, geschäftliche, finanzielle oder betriebswirtschaftliche Zwecke, sowie an geschäftsbezogene Überwachungs- oder Voraussagezwecke , DATA PROCESSING SYSTEMS OR METHODS, SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL, SUPERVISORY OR FORECASTING PURPOSES / G06F ELECTRIC DIGITAL DATA PROCESSING , Elektrische digitale Datenverarbeitung / G08G Anlagen zur Steuerung, Regelung oder Überwachung des Verkehrs , TRAFFIC CONTROL SYSTEMS |
Traffic flow prediction method based on time-space diagram attention neural network
European Patent Office | 2025
|Attention-based dynamic space-time diagram convolution traffic flow prediction system and method
European Patent Office | 2023
|Traffic flow speed prediction method based on attention space-time diagram convolutional network
European Patent Office | 2023
|European Patent Office | 2023
|European Patent Office | 2025
|