The Autoformer model is a deep learning architecture designed for long-term time series forecasting that enhances standard Transformer frameworks to better capture complex temporal patterns. Rather than relying on traditional point-wise self-attention mechanisms, it utilizes an auto-correlation mechanism that identifies dependencies and aggregates representations at the sub-series level based on inherent periodicity. To handle intricate temporal variations over extended prediction horizons, the architecture integrates progressive series decomposition blocks directly into its deep layers, continuously separating sequential data into distinct trend-cyclical and seasonal components for more accurate and computationally efficient multi-step predictions.
