Can a biLSTM be used to do closed loop forecasting?

I have read many papers talking about how biLSTM networks do better with time series forcasting compared to standard LSTM's. However, when I try to implament this in a closed-loop forecasting model, the network gets stuck and outputs the same time step over and over. What's weird is during the training the RMSE and loss for both training and validation are much lower than when an LSTM is used on the same data. This happens even when using this example: https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html and simply changing the LSTM layer to biLSTM. I'm not sure if I'm missing something or if it just can't be implamented in Matlab.

Answer 1

[Aldo]

Hi Davey, could you solve this problem? I have the same issue, and I can't fix it yet.