Using LSTM network in Nonlinear MPC design?

Hello everyone,

I would like to identify a system that has three inputs [u_1(k) u_2(k) y(k-1)] and single output as y(k) using LSTM time series estimation. I have a couple of questions regarding the implementation of this model in nonlinear MPC.

Based on the documentation of NMPC, I need to define a function for a state called StateFcn and an output called OutputFcn. As my model is based on the LSTM network, I was wondering how I can do that? Unfortunately, I couldn't find any example when I dig more into it. It worth mentioning that I am using GT-suit co-simulation as a virtual test machine, and I am going to implement this LSTM-based MPC to that.

Thank you in advance for your help.

[Ze Liu]

Hi Armin

Have you solved this problem? If so, can you please share the solution?

You can contact me via email:[e-mail]

Thank you very much!

[Dun-Ren Liu]

I face the same problem.can you please share the solution?

thanks~

[MD RAHAT]

I am having exactly same problem. Can you please guide me a little if you have found the solution to it

Answer 1

[Niccolò Dal Santo]

Hi Armin,

If I understand correctly you'd want train an LSTM for a time series with feedback. You can follow this example which shows how do that:

https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html?lang=en

You should define your inputs as a three-elements vector ([u_1(k) u_2(k) y(k)], hence numFeatures = 3), one response and train your LSTM accordingly.

For further reading, here is an example for training an LSTM with more than one input feature: https://www.mathworks.com/help/deeplearning/ug/sequence-to-sequence-regression-using-deep-learning.html

Hope this helps.

Cheers,

Niccolò

[Armin Norouzi]

Thank you for your response. I modeled my system using LSTM, and my main problem is how to use this model inside nlmpc mode. I understand that this model predicts sequence output for given sequence input. However, in nlmpc, I need to provide x(k+1) = f(u(k)) model, i.e., for given inputs in the previous time step, a model needs to be capable of estimating the next time step output. I would appreciate it if you could share your thoughts about this matter.