Experimental validation of deep Koopman MPC for real-time pasteurization unit control

This paper presents a novel deep Koopman Model Predictive Control framework for energy-intensive processes, addressing the critical challenge of real-time nonlinear control in pasteurization systems. The proposed approach employs neural networks to learn lifting functions that transform nonlinear system dynamics into a linear representation in the lifted space, enabling computationally efficient MPC implementation. A key innovation is the introduction of a lifted state correction mechanism that compensates for the mismatch between unmeasurable lifted states and real measurements, significantly improving control accuracy and practical implementability. Experimental validation on a laboratory-scale pasteurization unit demonstrates that the proposed deep Koopman MPC framework achieves 30 % improvement in control performance compared to conventional subspace identification methods, while maintaining real-time execution within 10ms on standard hardware. The enhanced control accuracy translates to reduced product waste and improved energy efficiency in pasteurization operations without additional computational overhead. This work represents one of the first experimental implementations of Koopman MPC in chemical process control and energy-intensive applications, establishing its practical feasibility for industrial deployment.