Effective MPC strategies using deep learning methods for control of nonlinear system

Model predictive control (MPC) is one of the important techniques for control of nonlinear and multivariable systems within constraints. Though the MPC ensures superior performance, it demands high computational resources to solve online optimization problems. With the advent of sophisticated deep l...

Full description

Saved in:
Bibliographic Details
Published in:International journal of dynamics and control 2024-10, Vol.12 (10), p.3694-3707
Main Authors: Rajasekhar, N., Nagappan, K. Kumaran, Radhakrishnan, T. K., Samsudeen, N.
Format: Article
Language:English
Subjects:
Automatic control
Closed loops
Complexity
Control
Control and Systems Theory
Control systems
Control theory
Deep learning
Dynamical Systems
Effectiveness
Engineering
Multivariable control
Neural networks
Nonlinear control
Nonlinear systems
Optimization
Performance evaluation
Performance measurement
Performance prediction
Predictive control
Recurrent neural networks
Servocontrol
Vibration
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Model predictive control (MPC) is one of the important techniques for control of nonlinear and multivariable systems within constraints. Though the MPC ensures superior performance, it demands high computational resources to solve online optimization problems. With the advent of sophisticated deep learning methods, neural networks can be employed to improve the computational efficiency of the MPC. In the present study, recurrent neural network (RNN) is used to approximate the MPC control law and tested on the benchmark multivariable quadruple tank (QT) system which consists of four interconnected tanks which demonstrate nonlinear and non-minimum phase characteristics. Modern variants of RNN, namely long short-term memory and gated recurrent unit, are used in this study to mimic the nonlinear MPC (NMPC) by learning the closed-loop simulation data. The optimum network architecture is chosen by altering the number of hidden nodes and layers till the required control performance on test dataset is reached. In order to test the efficacy of RNN as MPC controller, its performance is compared with the performance of linear MPC (LMPC). The automatic control and dynamic optimization (ACADO) toolkit is used in the optimization step of the NMPC computations. The servo and regulatory responses of the RNN-based MPC are compared with LMPC and evaluated in terms of standard control performance metrics such as integral squared error (ISE) and control effort (CE).
ISSN:2195-268X
2195-2698
DOI:10.1007/s40435-024-01426-3