[4E2-GS-2] Machine learning

[4E2-GS-2-04] Synthetic Data Augmentation for Time Series Forecasting

〇Kasumi Ohno¹, Kohei Makino¹, Makoto Miwa¹, Yutaka Sasaki¹ (1. Toyota Technological Institute)

Keywords:Time Series Forecasting, Data Augmentation

Recent years, deep learning models have achieved high prediction performance in Time Series Forecasting (TSF), which predicts future sequences from a given sequence. Training deep learning models requires a large amount of data; it is, however, difficult to prepare enough data since it takes time to collect the data. In this study, we augment the training data by using a variety of synthetic waveforms created with functions and other methods for data expansion. In the experiments, we employed Neural Laplace, which models the dynamics of time series in the Laplace domain, as a model for TSF and evaluated the effect of data augmentation with the synthetic waveforms on the Electricity Transformer Temperature (ETT) m2 dataset, a standard benchmark for TSF. We found that the data augmentation with the synthetic waveforms is effective for TSF on the ETTm2 dataset.