Traffic prediction method based on enhanced space-time diagram neural network

Abstract

The invention provides a traffic prediction method based on an enhanced space-time diagram neural network, and the method comprises the steps: modeling the time correlation and spatial correlation ofa road network based on a traffic prediction framework from a sequence to a sequence model, and constructing a directed weighted graph for the whole road network according to the upstream and downstream relationship of the road network; spatial correlation of a road network is captured through a diffusion graph convolutional network, spatial correlation characteristics of the road network are extracted, a time sequence with the spatial correlation characteristics is input into a recurrent neural network to capture time correlation of the road network, and then a prediction result is optimizedin the decoding process by an actor-critic algorithm in reinforcement learning; regarding A road network relation topological graph captured by each time slice as an actor in an intelligent agent anda recurrent neural network as a random strategy of a next action selected by the actor, judging the action selected by the actor by using critic, feeding back a dominance function, and enabling the actor to update strategy parameters according to the fed-back dominance function, so that prediction precision is greatly improved compared with a traditional method.

Description

technical field [0001] The invention belongs to the technical field of traffic forecasting, in particular to a traffic forecasting method based on a reinforced spatio-temporal graph neural network. Background technique [0002] With the rapid development of intelligent transportation systems, traffic forecasting has attracted more and more attention. It is an important part of traffic management system and an important part of traffic planning, traffic management and traffic control. Traffic forecasting can not only provide a scientific basis for traffic managers to perceive traffic congestion in advance and restrict vehicles, but also help travelers choose appropriate travel routes, thereby improving travel efficiency. However, complex spatio-temporal correlations in road networks complicate traffic prediction. [0003] There are many existing traffic prediction methods, the sensors used on the road include loop coil vehicle detectors, video vehicle detectors, infrared sen...

AI Technical Summary

Problems solved by technology

However, the training loss of the classic Seq2Seq model is supervised by the real samples collected by the sensor during the training process. However, there are no real samples at the time of testing, and the decoder completely relies on the generated output of the model itself to predict the next output, which will cause the model in There are differences in the next actions generated during training and testing, and the errors generated during this process will continue to accumulate. This difference is also called Exposure Bias.

Although methods such as planned sampling and confrontation generation network have also appeared to avoid exposure errors, these methods have some disadvantages, so effectively avoiding exposure errors is very necessary to improve prediction accuracy

Images