Traffic flow long-time prediction method and system based on improved Transform model

The invention discloses a traffic flow long-time prediction method and system based on an improved Transform model. The system comprises a data preprocessing module, a space-time traffic flow long-time prediction model construction module and an online prediction module. A time-space traffic flow long-time prediction model is constructed, the model adopts a multi-layer codec structure, and each layer of an encoder comprises a convolution attention module and a graph attention module which are used for modeling nonlinear time correlation and dynamic space correlation of traffic data; each layer of the decoder performs feature conversion on historical data by using a cross attention mechanism; after the encoder encodes historical traffic data, the data are input to the cross attention module of each layer of the decoder, judgment is made by referring to the historical data through a cross attention mechanism, historical traffic features are converted into future representations, a direct relation between histor