Predict a signal considering other related signal using RNN (LSTM)

Suppose we have a speed signal as a target in the regression problem.
Having several related signals such as acceleration, etc., I want to train a recurrent neural network that predicts the speed signal by taking input of features (acceleration signals, etc.).

Now I have three questions:

  1. Is it reasonable to use RNNs like LSTM for such a problem? And in general, what networks do you recommend for such a task?

  2. What is the best way to feed data to the model? Is it fine to give a Data-Frame where the features are in the columns and the rows represent the samples in time, as the input of the network?
    And how should hidden state and cell state be defined and update in the model so that the dependence of features and targets on the previous time step is considered.

  3. In the sliding window used in RNN, does the model use previously predicted values ​​to predict new values ​​or does it use the actual value in the dataset?

I will be happy to hear your comments.