MS-GATformer:Multi-Scale Hybrid Prediction Based on GAT-iTransformer

In multivariate time series, data at any time is often not only correlated with historical values, but also with the values of the other parameters. However, most of the existing methods are one-sided, focusing only on the correlation in time dimension or between parameters when processing multivariate time series, and unable to fully extract features of time series data. In addition, for the time series data under different observation scales, the trend and periodic change also show different states, reflecting the internal characteristics of the data itself. It is important to extract the intrinsic features of time series data at different observation scales. This paper proposes to use multi-window pooling to decompose the trend and period of complex time series data, then use convolution layers of different sizes to obtain the trend and period components of different scales, and finally extract the trend and period components of different scales based on the spatiotemporal feature extraction module integrated by graph attention neural network and iTransformer. To obtain the time series trend and periodic change characteristics at different granularities, the effect of the proposed model is demonstrated through some data sets.