Tune LSTM to get more accuracy on predicted input

Question

I am working on Machine Learning and faced this similar problem couple of time while using predicted result as an input to the model. It generally stop varying. PS: I am using L