TL;DR
WFTNet is a new architecture for long-term time series forecasting that combines the strengths of Fourier and Wavelet transforms. By using a novel Periodicity-Weighted Coefficient (PWC), the model adaptively balances global periodic patterns with local, fine-grained frequency variations, consistently outperforming current state-of-the-art models like TimesNet and FEDformer.
What Problem Does This Paper Address?
Predicting the long-term future of complex signals—like traffic flow, electricity consumption, or weather—is notoriously difficult. These signals often contain "periodicity," or patterns that repeat over time.
Current top-performing models (like TimesNet) use the Fourier Transform to identify these patterns. However, the Fourier Transform has a major limitation: it is a global operator. It excels at finding a steady heartbeat in a signal but struggles to capture "local" changes, such as a frequency shift that only happens for a short duration or non-stationary behaviors where patterns evolve. The authors argue that ignoring these local frequency structures leads to suboptimal forecasting.
Methods
The core innovation of WFTNet is the WFTBlock, which processes time series data through two parallel branches to capture a comprehensive "temporal-frequency" view.
- The Fourier Branch: This branch uses the Fast Fourier Transform (FFT) to identify the most dominant global frequencies. Following the approach of TimesNet, it reshapes 1D time series into 2D tensors based on these periods, allowing the model to use 2D convolutions to analyze intra-period and inter-period variations.
- The Wavelet Branch: Recognizing that Fourier misses local context, this branch employs the Continuous Wavelet Transform (CWT). Wavelets are mathematically designed to provide high resolution in both time and frequency domains, making them perfect for spotting transient or localized periodic structures.
- Periodicity-Weighted Coefficient (PWC): This is the "brain" of the fusion stage. The PWC ($\alpha$) is a calculated metric that measures the relative strength of global periodicity in a signal.
- If a signal is highly periodic (like electricity usage), $\alpha$ is high, and the model relies more on the Fourier branch.
- If the signal is irregular or non-stationary, $\alpha$ is low, and the model leans on the Wavelet branch.
Datasets
The authors evaluated WFTNet on seven standard benchmark datasets for long-term forecasting:
- ETT (Electricity Transformer Temperature): Four sub-datasets (ETTh1, ETTh2, ETTm1, ETTm2) recorded at different granularities.
- Traffic: Hourly road occupancy rates.
- ECL (Electricity Consumption Load): Hourly electricity consumption of 321 clients.
- Weather: Local climatological data containing 21 variables.
Key Insights and Ideas
- Global vs. Local Synergy: The paper demonstrates that global and local periodicities are not mutually exclusive; they are complementary. A model that only looks at one (like TimesNet with Fourier) is essentially "blind" to half the story.
- Adaptive Fusion is Key: Different datasets have vastly different periodic characteristics. As shown in Figure 2, the ECL dataset has very clear, strong periodicity, while the ETTh2 dataset is much noisier. The PWC allows WFTNet to change its strategy dynamically for each dataset—or even each channel within a dataset—without manual tuning.
Key Results
WFTNet achieved impressive results across the board, securing the best performance in 19 out of 32 tested scenarios (combinations of datasets and prediction horizons).
- Superiority over SOTA: WFTNet consistently beat TimesNet, which was the previous benchmark for frequency-based forecasting. For example, on the ECL dataset at a prediction length of 96, WFTNet achieved an MSE of 0.164, compared to TimesNet's 0.167.
- Ablation Success: The authors proved the necessity of both branches. In their ablation study (Table 2), the full WFTNet outperformed both "Fourier-Only" and "Wavelet-Only" versions, validating that the PWC-based fusion effectively harvests the best of both worlds.
Limitations and Future Work
While the paper presents a strong case for WFTNet, a few areas remain for exploration:
- Computational Overhead: The Continuous Wavelet Transform (CWT) can be computationally more expensive than the FFT. While the authors use efficient convolutional structures, the scaling of this approach to extremely high-dimensional or high-frequency data could be a bottleneck.
- Mother Wavelet Selection: The study primarily uses the Morlet wavelet. Exploring how different "mother wavelets" affect performance on specific types of time series (e.g., financial vs. physical sensors) could be a fruitful research direction.
- Hyperparameter Sensitivity: The fusion equation uses a hyperparameter $n$ to modulate the PWC. The paper does not discuss in depth how sensitive the model is to this specific value across different domains.
Summary updated