I have been involved with pytorch only recently, so I ask for some leniency with me.
I do not know if I answer correctly to my current project or choose the right approaches. I hope you can help me.
I have the data of a power generation plant as time series and would like to predict how some values will behave. To be precise I have 6 input parameters and 4 output parameters, the 4 output parameters I would like to predict some time steps into the future.
I tried to use LSTM to forecast my exit savings. Unfortunately I did not succeed here.
Can you give a few tips or possibly provide prommier approaches.
I have spent some time in the last few days, unfortunately I have to say I still do not quite understand how I can use it with 4 input parameters and 6 output parameters.
If you don’t have a feedback feom output to next input, typically, you’d have a Linear (4, hidden_size) an RNN/GRU/LSTM with hidden_size input/output of sort and a Linear(hidden_size, 6).
Depending on the nature of the data, there are customerary models and/or preprocessing. For example Graves’ classic handwriting model uses offsets instead of coordinates to make the prediction targets more stationary and uses a Mixture Density Network to model the outputs. I have implemented the handwriting generation in PyTorch here, but I’d highly recommend reading the original paper.