Using autoencoder for time series prediction

I was recently reading a paper on time series prediction using deep learning methods. There I found a technique named "Variational Autoencoder" to predict time series data. I understand how LSTM and rolling methods could be used to forecast this type of data. However, when it comes to autoencoders, such networks simply encodes the data into a lower and compressed dimension. If we compress a time series features( here I mean, the timesteps) what benefit we get from there? Also how this method could be used to forecast future time series? I am talking about this paper:

https://www.sciencedirect.com/science/article/pii/S096007792030518X

Also, I have gone through this article on autoencoders:

https://blog.keras.io/building-autoencoders-in-keras.html