Time Series Forecasting Based on Structured Decomposition and Variational Autoencoder

Time series forecasting based on decomposition method usually decomposes a complex time series into some simple components, such as long-term and seasonal trends, which are more easy to be predicted. Though long-term and seasonal trends are vital in time series foreasting, it may be insufficient for those not having such obvious characters. This paper proposes a time series forecasting model named SD-VAE based on structured decomposition and variational autoencoder(VAE). The structured decomposition module decomposes a time series into long-term component, seasonal component, short-term component and co-evolving component, and the VAE module learns the representation of them, through which the future value of each decomposed component are predicted and then fused by a neural network. To get a better decoupled representation of each decomposed component, mutual information constraints are added in the latent space of VAE. Extensive experiments prove that our model performs the best in 3 out of 4 public datasets, with the other one ranking second, against the most representation learning and end-to-end forecasting models.