An Autoformer-CSA Approach for Long-Term Spectrum Prediction

In this letter, we develop an Autoformer with a series channel-spatial attention module (CSAM) (Autoformer-CSA) for long-term spectrum prediction. More specifically, the CSAM ingeniously replaces 2-dimensional (2D) convolution in image attention (including channel attention and spatial attention) with 1-dimensional (1D) convolution. The CSAM replaces Autoformer's feed-forward network. It is used to assign different concentrations to features mapped to the high-dimensional space, improving the learning ability of the Autoformer. We follow the series decomposition block and auto-correlation mechanism of the Autoformer. Experiments on a real-world dataset show that the Autoformer-CSA is superior to the state-of-the-art benchmarks.

Automated summary

In this letter, we propose an Autoformer with a series channel-spatial attention module (CSAM) (Autoformer-CSA) for long-term spectrum prediction. The CSAM creatively replaces 2D convolution with 1D convolution in image attention, significantly improving the learning ability of the Autoformer. Experimental results demonstrate that the Autoformer-CSA outperforms state-of-the-art benchmarks.