Space-time attention mechanism method for traffic prediction

Abstract

The invention provides a space-time attention mechanism method for traffic prediction, and belongs to the field of traffic prediction. The invention relates to a traffic prediction framework of an end-to-end solution, which can be used for modeling spatial, short-term and long-term periodic dependencies. The APTN first models spatial dependencies and periodic dependencies using an attention mechanism of an encoder. Our model can more easily capture these dependencies because each node is to process all other nodes in the network. Time attention is then applied to select a relevant encoder concealment state across all time steps. We use a traffic data set of the real world to evaluate the proposed model and observe consistency improvements on the most advanced baseline.

Description

technical field [0001] The invention belongs to the field of traffic prediction, and in particular relates to a space-time attention mechanism method for traffic prediction. Background technique [0002] The data of the traffic forecasting system has the characteristics of space and time (periodicity, time series). At this stage, the method of traffic prediction is mainly to extract spatial correlation through deep learning CNN, and RNN or its variant LSTM / GRU to model time dependence. [0003] These approaches to jointly model spatial, temporal, and external feature dependencies by integrating CNN and LSTM may not be able to adequately model spatial dependencies because convolution operations only aggregate local information and require many layers to learn long-range spatial dependencies. This problem is alleviated using multi-layer convolutions, which consider distance, functional similarity and transport connectivity when modeling spatial dependencies. But it needs to ...

AI Technical Summary

Problems solved by technology

[0003] These approaches to jointly model spatial, temporal, and external feature dependencies by integrating CNN and LSTM may not be able to adequately model spatial dependencies since convolution operations only aggregate local information and require many layers to learn long-range spatial dependencies

Images