Can MPCPy support external trained model

[Patrick Duhirwe Nzivugira]

Hello,

I have DNN models trained and saved using Keras. Can I use those models to represent my system behavior?

If possible, can you provide guidelines for it?

Thank you.

[David Blum]

Hello, 

Thank you for your question.  I'd like to understand it a little more.  Are you looking to use the DNN models to solve the optimal control problem? 

Currently, DNN models are not supported.  One way to utilize such a model would be to implement a new model class that extends mpcpy.models._Model and uses MPCPy exodata, units, and utility functionality.  The class mpcpy.models.Modelica could be used as a template.  If wanting to use this model in an optimal control problem, new optimization package type and problem type classes could be created that extend mpcpy.optimization._Package and mpcpy.optimization._Problem to utilize the new DNN model type, using for instance the classes mpcpy.optimization.JModelica and mpcpy.optimization.EnergyMin as templates for package type and problem type respectively.

Best Regards, 

Dave