Traffic flow prediction method and system based on multi-granularity progressive transfer learning

Abstract

The application discloses a traffic flow prediction method and system based on multi-granularity progressive transfer learning, and relates to the technical field of intelligent transportation. The method comprises the following steps: acquiring traffic flow time series of a source domain and a target domain and constructing a space-time graph structure; learning macroscopic traffic rules shared across cities through coarse-granularity transfer by data enhancement and contrastive learning; inputting the time series into a shared GRU encoder to extract time features, learning self-adaptive graph structures through respective graph learners, performing spatial feature aggregation through GAT, combining a domain adversarial network and contrastive learning to perform medium-granularity transfer, and learning domain-invariant functional area features; performing Fourier transform to extract frequency domain features, calculating sample similarity to generate a weight matrix, and performing fine-granularity transfer; using target domain labeled data to jointly fine-tune the model, and outputting a prediction result. Through multi-granularity progressive transfer learning, the application realizes cross-domain traffic flow accurate prediction in a data-scarce scenario.

Images