Using a loop to predict longer vectors with a LSTM

Question

I have a model based on LSTMs that can predict a vector output based on a vector input. I can't increase the size of the output because :

• I would need a larger network to obtain good results
• It would take more time to train
• The behaviour of my timeseries can be captured in the size of the vector I'm already predicting

So I tried to predict an output, and then use that prediction as a new input in my model and predict a new vector (without re-training). I iterate 5 times and I get this result :

As you can see, the first prediction is pretty good. Then, my model is lost and loses its accuracy. Do you know how to fix it ?

Thanks

Answer 1

I'm going to answer to myself since I managed to fix that issue. The problem is that my model was overfitting waaaay too much and I had to add a dropout between each layer.

Now the result is better :

I will now re-tune my model and improve it a bit.

note to myself : don't be lazy Julien, plot the validation loss next time

EDIT :

I trained it during 36 mn with more neurons per layer. Without overfitting, it's way way better.

Now I can predict 8 days of datapoints with a better accuracy.