LSTM convergence

Hello All!

I am working on the task of univariate time series prediction. I have 2 cases:

  1. First case, when I use already prepared test data (sequence of observations) to predict value at next time step.
  2. Second case, when I use predicted values as input to predict value at the next time step. E.g. I use [x1, x2, x3] to predict x4, at next time step I will use [x2, x3, x4] to predict x5, and so on. Values after x3 are prediction, not from test data.

For the first case everything is working fine, prediction are appropriate.

But when I use predicted values It does not converges. Despite first few values a close to real ones.

I checked several times the types of input and they are all the same.
Guys, I really need your help!