Multi steps time series prediction, with multi features

Element-Research / rnn

Recurrent Neural Network library for Torch7's nn

BSD 3-Clause "New" or "Revised" License

939 stars 313 forks source link

Multi steps time series prediction, with multi features #336

Open iwinterlovermy opened 8 years ago

iwinterlovermy commented 8 years ago

Hi there,

I'm currently trying to solve a time series problem using LSTM. I'm trying to predict 7 days ahead the amount of loading I will have. The amount of loading is affected by the suppliers supplying the goods to me. Assuming I have 10 suppliers. I have collected 180 days of data. LSTM requires the training set to be in the form of [rows, time steps, features]. In my case, do I prepare the data in the form of [180, 7, 10]?

Thanks in advance!

juesato commented 7 years ago

You should check out the docs for Sequencer - they should be what you want.

The input to a Sequencer module is a table of length number of time steps where each input output pair has shape batchSize * dimension. So if your inputs are vectors of length 10, and you're using a batch size of 1, then you want your outputs at each time step to be your 10 suppliers for that day, and your output to be loading 7 days in the future.