Time series prediction method based on multivariate long-term time series analysis

The invention discloses a time series prediction method based on multivariate long-term time series analysis. The invention provides a brand new high-performance time series prediction model to realize time series prediction. The model is composed of a statistical value linear fitting network, a multi-dimensional graph learning network, a dual-data flow learning network and a bias generation network. And the anti-overfitting capability of the model and the learning prediction capability of the time sequence information are improved by using a dual-data-stream learning mode. A multi-dimensional graph learning network is used for carrying out feature extraction in a multi-dimensional domain of time series data, and features are extracted in a multi-view mode in combination with a time domain double-data-stream network. A bias generation network is combined with a fitting statistical value to solve the problem of distribution drift on the basis of RevIN. Compared with advanced DLiner and TimesNet, the MVENet mod