A recurrent neural network-based approach for joint chance constrained stochastic optimal control. (August 2022)

• Record Type: Journal Article
• Authors: Yang, Shu-Bo; Li, Zukui; Moreira, Jesús
• Is Part Of: Journal of process control. Volume 116(2022)
• Page Start: 209
• Page End: 220
• Publication Date: 2022-08
• Subjects: Optimal control -- Recurrent neural network -- Stochastic model predictive control -- Joint chance-constrained optimization -- Surrogate modelling -- Sample average approximation; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.06.012
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: A recurrent neural network (RNN)-based approach is proposed in this paper to handle joint chance-constrained stochastic optimal control problems (SOCP) and stochastic model predictive control (SMPC) implementations. In the proposed approach, the joint chance constraint (JCC) in a SOCP is first reformulated as a quantile-based inequality. Then, the sample average approximation (SAA) method is used to build the RNN-based surrogate model for the quantile function. Afterwards, the RNN-based model is embedded into the probabilistic constraint of the SOCP. Subsequently, the SOCP involving the RNN-based model can be solved using a deterministic nonlinear optimization solver. Moreover, while applying the proposed approach to the SMPC, the SOCP involving the RNN-based model is solved repeatedly at different sampling instants, based on different initial system states. The proposed approach is applied to a numerical illustrating example and a chemical process case study to demonstrate its capability of handling the SOCP and the SMPC implementation. Highlights: Quantile reformulation of joint chance constraints and sample based empirical approximation. Approximation of quantile function through recurrent neural network based surrogate model. Stochastic optimal control problem is deterministically solvable though embedded RNN model. Easy extension and implementation of the proposed approach to stochastic model predictive control problem.