Abstract
Deep learning models for time-series forecasting often struggle with data scarcity, where limited or missing historical records hinder the learning of complex temporal dynamics. While data augmentation is a standard solution in other domains, existing time-series methods frequently fail to balance the need for sample diversity with the preservation of temporal coherence. This paper presents STAug (Spectral and Time Augmentation), a novel framework that generates synthetic data by operating in both the frequency and time domains. By utilizing Empirical Mode Decomposition (EMD) to manipulate intrinsic signal components and a modified Mix-up strategy to blend temporal sequences, STAug significantly improves forecasting accuracy. Experimental results across multiple benchmarks demonstrate that STAug reduces Mean Square Error (MSE) by an average of 28.2% and exhibits particular robustness in low-data regimes.
Introduction
Time-series forecasting is a critical component of modern infrastructure, powering everything from energy grid management to financial market analysis. However, the performance of state-of-the-art deep learning models is heavily dependent on the availability of large, high-quality datasets. In practice, real-world sensor data is often characterized by "data scarcity" due to sensor failures, short collection windows, or high costs of data acquisition.
Traditional data augmentation techniques, such as those used in computer vision (e.g., rotation or cropping), do not translate easily to time-series forecasting. In classification tasks, the primary goal is label preservation. In forecasting, however, the model must capture the underlying physical process and temporal dependencies. Existing methods like "permutation" (shuffling segments) often destroy the temporal order, while "filtering" (noise removal) can lead to a lack of diversity in the training set. STAug addresses this by ensuring that synthetic samples are both diverse enough to aid generalization and coherent enough to reflect realistic physical dynamics.
The STAug Framework
STAug operates through a dual-domain approach, targeting both the spectral characteristics and the raw temporal sequences of the data.
1. Frequency-Domain (Spectral) Augmentation
Standard frequency-domain methods often rely on the Fourier Transform, which assumes stationarity—a property rarely found in real-world time series. STAug instead employs Empirical Mode Decomposition (EMD). EMD decomposes a complex signal into a finite set of Intrinsic Mode Functions (IMFs) and a residual trend. These IMFs represent distinct oscillatory modes, ranging from high-frequency noise to low-frequency seasonal patterns.
To generate an augmented sample, STAug reassembles these IMFs using random weights. This process preserves the "ingredients" of the original signal (maintaining coherence) while varying their relative importance (providing diversity). This principled decomposition allows the model to encounter variations of the signal that are physically plausible but structurally distinct from the original training samples.
2. Time-Domain Augmentation
Following spectral augmentation, STAug applies a modified Mix-up strategy. This involves taking two randomly sampled, spectrally augmented series and performing linear interpolation between them. Crucially, this interpolation is applied to both the historical input window and the future ground-truth labels. This creates "in-between" samples that represent plausible transitions between known data points, further smoothing the model's decision boundaries and improving its ability to interpolate between observed states.
Experimental Evaluation
Datasets and Setup
The researchers evaluated STAug using five multivariate benchmarks widely recognized in the forecasting community:
- ETTh1 & ETTh2: Hourly electricity transformer temperature data.
- ETTm1 & ETTm2: 15-minute interval electricity transformer temperature data.
- Exchange: Daily currency exchange rates for eight countries.
The framework was integrated with the Informer architecture, a state-of-the-art transformer-based forecaster. Experiments utilized a standard 7:2:1 split for training, validation, and testing, with a context window of 96 time steps used to predict horizons ranging from 96 to 720 steps.
Key Results
STAug demonstrated superior performance compared to the base Informer model and seven other competitive augmentation baselines, including GAN-based and STL-based (Seasonal-Trend decomposition using LOESS) approaches.
- Error Reduction: STAug achieved an average reduction in Mean Square Error (MSE) of 28.2% and a reduction in Mean Absolute Error (MAE) of 18.0% relative to the best-performing baselines.
- Statistical Significance: The improvements were validated using Friedman and Wilcoxon-signed rank tests, confirming the results were statistically significant (p < 0.05).
- Robustness to Scarcity: A critical finding was STAug's performance in "cold-start" scenarios. When training data was reduced to only 10% of the original set, STAug provided the most substantial performance gains, highlighting its utility for emerging applications where historical data is limited.
Discussion and Insights
The success of STAug stems from the joint consideration of time and frequency domains. Ablation studies confirmed that removing either the EMD-based spectral augmentation or the time-domain Mix-up led to a noticeable decline in performance. While the time domain captures localized transitions, the frequency domain captures the broader periodicities essential for long-term forecasting.
Qualitatively, models trained with STAug produce forecasts that more accurately track the "peaks and valleys" of the ground truth. By training on a more diverse set of coherent oscillations, the model learns to better distinguish between meaningful signal variations and random noise.
Conclusion and Future Work
STAug provides a robust framework for time-series data augmentation by balancing diversity and coherence. While the current implementation uses EMD, which can be more computationally intensive than the Fast Fourier Transform (FFT), the components can be pre-computed to mitigate overhead. Future research directions include developing automated selection procedures to tune augmentation parameters (such as the Mix-up alpha) dynamically based on dataset characteristics, and exploring the framework's efficacy in highly stochastic domains like high-frequency financial trading.
Summary updated