[4Xin2-63] Data Augmentation for Long-term Time Series Forecasting with High-quality Time Series Generation
Keywords:Long-term Time Series Forecasting, Time Series Generation, Pre-training
This study focuses on data augmentation in Long-term Time Series Forecasting (LTSF), which is the prediction task of long-term future series from a given data.
We propose a novel time-series generation model, iTransGAN, that can handle long-term time series and investigate the effect of data augmentation using a large amount of synthetic data generated by the model. In the experiments, we evaluate the quality of the synthetic data and the effect of data augmentation using the synthetic data with the Electricity Transformer Temperature (ETT) dataset, one of the standard benchmark datasets for LTSF.
The results show that the proposed iTransGAN model can generate higher-quality long-term time-series data than existing methods. We also confirmed that pre-training with the synthetic data provides higher prediction accuracy than learning from scratch. In addition, we found that pre-training with synthetic data enables the learning of large-scale models even when the models learn well only with scarce real data.
We propose a novel time-series generation model, iTransGAN, that can handle long-term time series and investigate the effect of data augmentation using a large amount of synthetic data generated by the model. In the experiments, we evaluate the quality of the synthetic data and the effect of data augmentation using the synthetic data with the Electricity Transformer Temperature (ETT) dataset, one of the standard benchmark datasets for LTSF.
The results show that the proposed iTransGAN model can generate higher-quality long-term time-series data than existing methods. We also confirmed that pre-training with the synthetic data provides higher prediction accuracy than learning from scratch. In addition, we found that pre-training with synthetic data enables the learning of large-scale models even when the models learn well only with scarce real data.
Please log in with your participant account.
» Participant Log In