Using stochastic programming to train neural network approximation of nonlinear MPC laws

To facilitate the real-time implementation of nonlinear model predictive control (NMPC), this paper proposes a deep learning-based NMPC scheme, in which the NMPC law is approximated via a deep neural network (DNN). To optimize the DNN controller, a novel “optimize and train” architecture is designed...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Automatica (Oxford) Ročník 146; s. 110665
Hlavní autori: Li, Yun, Hua, Kaixun, Cao, Yankai
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier Ltd 01.12.2022
Predmet:
Deep neural networks
Model predictive control
Nonlinear systems
Parallel computation
Policy learning
Stochastic optimization
Model predictive control
Parallel computation
Deep neural networks
Stochastic optimization
Policy learning
Nonlinear systems
ISSN:0005-1098, 1873-2836
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:To facilitate the real-time implementation of nonlinear model predictive control (NMPC), this paper proposes a deep learning-based NMPC scheme, in which the NMPC law is approximated via a deep neural network (DNN). To optimize the DNN controller, a novel “optimize and train” architecture is designed, where the processes of data generation and neural network training are combined together to result in a single large-scale stochastic optimization problem. Unlike the conventional “optimize then train” approach, our proposed one directly optimizes the closed-loop performance of the DNN controller over a finite horizon for a number of initial states. The important features of our proposed scheme are that it can deal with set-valued optimal MPC input, and a probabilistic guarantee of constraint satisfaction can be concluded for the closed-loop system without simulating the DNN controller. With our proposed scheme, an increased number of training scenarios leads to improved constraint satisfaction of the derived DNN controller, which is not necessarily true for the “optimize then train” approach. Statistical approaches for validating closed-loop control performance are also discussed. Furthermore, computational methods are introduced to efficiently solve the resulting stochastic optimization problem. The effectiveness of the proposed scheme is extensively illustrated with several numerical simulations. Compared with the conventional “optimize then train” approach, our proposed approach exhibits better closed-loop constraint satisfaction for all considered case studies.
ISSN:0005-1098
1873-2836
DOI:10.1016/j.automatica.2022.110665