Control Lyapunov barrier function-based predictive control of nonlinear systems using physics-informed recurrent neural networks

Control Lyapunov-barrier functions (CLBF) have been effectively employed in model predictive control (MPC) to ensure both closed-loop stability and operational safety in input-constrained nonlinear systems. In this work, we propose a novel CLBF-MPC framework that leverages physics-informed partially-connected recurrent neural network (PCRNN) models to enhance prediction accuracy by incorporating a priori process structural knowledge. The PCRNN architecture, designed based on known process interconnections, enables improved approximation of nonlinear dynamics which, when incorporated into a CLBF-MPC, allows for improved process operational safety by avoidance of unsafe regions in the state-space that would normally be encountered under regular MPC operation. The effectiveness of the proposed PCRNN-based CLBF-MPC is demonstrated through application to a chemical process example, where it achieves superior predictive performance and successfully maintains system safety by fully avoiding the bounded unsafe region, unlike the fully-connected black-box RNN model when incorporated into the same CLBF-MPC.