Daformer: A Novel Dimension-Augmented Transformer Framework for Multivariate Time Series Forecasting

Recently, many deep learning-based models have been proposed for multivariate time series forecasting (MTSF). However, these models usually fail to fully exploit two crucial features of multivariate time series: periodic characteristics and cross-variable dependencies. To address these two limitations, we present a novel Dimension-augmented Transformer framework named Daformer, which includes the Dimension-augmented Module, the Feature Fusion Module, and the Cross-variable Transformer Encoder (CTE) block. The Dimension-augmented Module and the Feature Fusion Module are used to extract the periodic distribution properties of the time series variables, while the CTE block is used to extract the cross-variable dependencies. Experimentally, Daformer achieves consistent state-of-the-art performance in six real-world forecasting benchmarks, reducing the mean squared error by an average of 3.4%, and the number of model parameters by 67%. Our code is publicly accessible at https://github.com/qiuyueli123/Daformer .