Data Augmentation for Short-Term Time Series Prediction with Deep Learning

In this paper, a hybrid data augmentation technique for short-term time series prediction is proposed in order to overcome the underfitting problem in deep learning models based on recurrent neural networks such as Long Short-Term Memory(LSTM) and Gated Recurrent Unit (GRU). The proposal hybrid technique consists of the combination of two basic data augmentation techniques that are generally used for time series classification, these are: time-warping and jittering. Time-warping allows the generation of synthetic data between each pair of values in the time series, extending its length, while jittering allows the synthetic data generated to be non-linear. To evaluate the proposal technique, it's experimented with three non-seasonal short-term time series of Peru: CO2 emissions per capita, renewable energy consumption and Covid-19 positive cases, it is considered that predicting non-seasonal time series is more difficult than seasonal ones. The results show that the regression models based on recurrent neural networks using the selected time series with data augmentation improve results between 16.318% and 42.1426%.