Model Predictive Control based on Long-Term Memory neural network model inversion

Abstract

Long Short-Term Memory (LSTM) neural networks are well suited for representing time series as, compared to other neural networks, their structure avoids vanishing or exploding gradients. LSTM has been embedded into Model Predictive Control algorithms in order to forecast the behavior of nonlinear systems. The new algorithm presented in the paper is of a different nature, as the LSTM network approximates the inverse of the system over a receding horizon and provides a sequence of future inputs as a function of a specified output trajectory. The main advantage of the method appears when the desired output trajectory is generated from a small set of parameters, for example, a convergence rate. The Model Predictive control optimizes its criterion with respect to this small set of variables, and the LSTM supplies the corresponding future control inputs. Eventually, the modeling error of the LSTM can be compensated by feeding the control sequence to the forward model and updating the controller according to the output deviation. The algorithm allows to design Model Predictive controllers for nonlinear systems in a generic way, using a very small number of decision variables even with a long receding horizon.