A Pyramid Attention-Based Model for Joint Time-Frequency Forecasting

Abstract

Long-horizon forecasting is important for applications such as electricity-load management and climate analysis, but a model must represent both distant temporal interactions and local periodic variations to perform well. We present RDDPformer, a two-path Transformer that processes temporal patterns with dilated convolutions and spectral patterns with Fourier decomposition. Instead of converting the spectrum to magnitude and phase, the proposed model feeds its real and imaginary components, obtained through the DFT, into separate frequency encoders. A convolution-enhanced pyramid attention module extracts local periodic structure, while cross-attention selects frequency information conditioned on temporal features. RevIN is applied to reduce the influence of distribution changes. On ETTh1, ETTh2, ETTm1, ETTm2, and Weather, RDDPformer lowers the MSE relative to Pyraformer by 42.7%–89.9% across the evaluated horizons. The ablation results further show that both the real/imaginary representation and the cross-attention module make complementary contributions.