Efficient operation strategy for the thermal plants via hierarchical MPC based on reduced-order Koopman modeling

Operational efficiency has become a paramount concern in modern thermal power plants, where improved control directly translates to significant resource conservation. To address the inherent strong nonlinearities in boiler-turbine systems, this study firstly develops a novel hierarchical model predictive control (HMPC) framework based on an advanced reduced-order Koopman modeling approach. The proposed methodology combines sparse neural networks (NN) with the Generalized Sparse Identification of Nonlinear Dynamics (GSINDy) algorithm to transform nonlinear dynamics into a more tractable linear representation, while employing weighted proper orthogonal decomposition (POD) for effective dimension reduction. Compared to conventional Extended Dynamic Mode Decomposition (EDMD)-based Koopman models, our approach achieves superior prediction accuracy without substantially increasing model complexity. The resulting HMPC framework implements a two-layer control architecture: the high-layer computes optimal trajectories over extended horizons, while the low-layer ensures precise tracking through shorter intervals, with seamless coordination enabled by an innovative waysets-based communication mechanism. Comprehensive simulation studies under various load-varying scenarios demonstrate that the proposed framework outperforms traditional centralized tracking MPC in both control efficiency and dynamic response. It meets the requirement of control efficiency under load demand variations.