Increase autoencoder performance

I am trying to implement an autoencoder technique with LSTM layers. And then extract features and use them to train a predictive model. Autoencoder performs badly, I am looking for help to improve it.