FEDformer: Frequency Enhanced Decomposed Transformer
FEDformer (Frequency Enhanced Decomposed Transformer) · Also known as: Frequency Enhanced Decomposed Transformer, FED-Transformer, Frequency Domain Transformer, Frekans Tabanlı Ayrıştırılmış Dönüştürücü
FEDformer is a Transformer-based architecture for long-term multivariate time-series forecasting, introduced by Zhou et al. at ICML 2022. Its core innovation is the combination of seasonal-trend decomposition with frequency-domain attention: instead of computing full token-to-token attention in the time domain, FEDformer projects queries, keys, and values into the frequency domain via Fourier or wavelet transforms and operates on a randomly selected subset of frequency components, achieving linear complexity while preserving global temporal structure.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
FEDformer is well suited for long-horizon multivariate time-series forecasting tasks—typically predicting hundreds to thousands of steps ahead—where the data exhibits clear periodic or seasonal patterns (e.g., energy consumption, weather, traffic, finance). It assumes stationarity at the frequency level and benefits most when dominant spectral components carry predictive information. For very short horizons, simpler linear baselines may suffice. For purely non-periodic or highly irregular series, frequency-domain assumptions may not hold, making models like Temporal Fusion Transformer or N-BEATS better alternatives.
Strengths & limitations
- Linear O(L) time and memory complexity via sparse frequency sampling, enabling long-context forecasting
- Explicit seasonal-trend decomposition provides interpretable model components and improves generalization
- Frequency-domain operations capture global periodic structures that local time-domain attention often misses
- Supports both Fourier and wavelet variants, offering flexibility for different signal characteristics
- Random frequency mode selection introduces stochasticity; different random seeds can yield slightly different results
- The model architecture is relatively complex, with multiple decomposition blocks and specialized attention layers that increase implementation overhead
- Assumes that a small set of frequency components dominates the series; this assumption fails for non-stationary or chaotic signals
- Subsequent work (e.g., PatchTST, DLinear) has shown that much simpler architectures can be competitive on standard benchmarks, questioning FEDformer's architectural complexity
Frequently asked
How does FEDformer differ from Autoformer?
Both models use seasonal-trend decomposition and achieve linear complexity, but Autoformer computes auto-correlation in the time domain using FFT as a computational shortcut. FEDformer explicitly moves the attention mechanism into the frequency domain and learns on a sparse subset of frequency modes, making frequency-domain processing a first-class architectural choice rather than a computational trick.
Why is random frequency mode selection used instead of always selecting the top-K modes?
The authors found that randomly sampling frequency components acts as implicit regularization and improves generalization. Deterministically selecting the highest-magnitude modes can overfit to dominant periodic patterns in the training data, whereas random selection forces the model to learn more robust representations across the full spectral distribution.
Can FEDformer handle univariate forecasting, or is it designed only for multivariate data?
FEDformer can handle both univariate and multivariate settings. In univariate mode each channel is processed independently by the same architecture. In multivariate mode, the encoder and decoder jointly process all channels, allowing the model to capture inter-variable correlations through the shared frequency-domain representations and cross-attention mechanism.
Sources
- Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., & Jin, R. (2022). FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. ICML. link ↗
How to cite this page
ScholarGate. (2026, June 2). FEDformer (Frequency Enhanced Decomposed Transformer). ScholarGate. https://scholargate.app/en/deep-learning/fedformer
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →