Decomposition transformers are deep learning architectures designed for time series analysis and forecasting that embed series decomposition mechanisms directly into the internal layers of a transformer model. Rather than treating time series decomposition as a separate data preprocessing step, these models integrate decomposition blocks throughout their encoder and decoder structures to progressively separate complex temporal sequences into distinct components, such as long-term trend-cyclical patterns and periodic seasonal variations. By isolating these individual dynamics within intermediate representations, decomposition transformers simplify intricate temporal patterns, reduce noise, and improve the efficiency and accuracy of learning long-range dependencies across complex sequential data.
