An Analysis of Combined Data Augmentation in Time Series Prediction Tasks Using Discrete Wavelet Multilevel Decomposition

In time series prediction, the generalization error of a model is expected to decrease as the size of its training set increases. However, limited availability of samples is often a challenge in many cases. This study proposes a data augmentation strategy for time series prediction that can be applied with various augmentation algorithms. The approach employs a decomposition strategy based on discrete wavelet transform, specifically targeting the multi-layer high-frequency subseries obtained from wavelet decomposition. The aim is to generate more informative features that better align with real-world data. Through extensive experimentation, the optimal hybrid algorithm for the two high-frequency subseries layers is determined, and two different knowledge fusion methods are employed to process the augmented samples along with the original samples, achieving the best fusion effect. The impact of different decoder architectures in the Transformer model on prediction accuracy is also discussed. The effectiveness of the proposed data augmentation strategy and the advancements of the Transformer model are validated through experiments conducted on the eight countries most severely affected by the pandemic.