Signal prediction with LSTM(deep learning)


As part of my research I am trying to familiarize myself with LSTM neural networks. These neurons can be used to predict the following data in a time series.

I first made a small network and a training time series which is simply the cosine function. In this case the network manages to find the new values (valid prediction), see image below with in red the training base and in blue the prediction.

I’m satisfied with this result but to go further I tried to have a training base with a double period, but in this situation my LSTM network can’t predict correctly; I tried to change the parameters or the architecture but I really can’t reproduce this double period in prediction. Do you know how it is possible to do this? I put below an example of this double period and the best prediction I could get. The first 10 000 values represent the database, the values following the prediction.

the image i was talking about:

Hello there,

Sounds like you need to tune your response time-step, and/or you are not standardising the data correctly.

Hope this helps