Cascaded machine learning-based neural Model Predictive Control for real-time thermal management of a sCO2 serial coaxial all-glass evacuated solar collector

There is a growing imperative for real-time thermal energy management in solar collectors to ensure system stability and facilitate integration with advanced technologies, such as Carnot batteries. To address this need, the present study proposes a Cascading Feedforward Neural Network (CFNN)–Model Predictive Control (MPC) framework, which enables the prediction and control of complex nonlinear thermal behavior under soft constraints on useful heat gain and exergy gain. A preliminary evaluation of a series-configured, all-glass coaxial evacuated tube collector using supercritical CO 2 was performed. A parametric study demonstrates that the inlet temperature, whether below or above the peak isobaric specific heat, significantly dictates the outlet thermal behavior and exergy efficiency, ultimately shifting the system's optimal operating point. Furthermore, parametric evaluation revealed that the system, utilizing the coaxial fluid conduit, reaches a maximum temperature of 308.15 °C and an exergy efficiency of 18.87%. Consequently, the CFNN, by virtue of its inherent cascading architecture, demonstrated superior predictive accuracy, yielding an R 2 value of 0.99948. The CFNN–MPC framework optimizes real-time thermal management by balancing outlet temperature regulation with minimal exergy destruction. While disturbances such as dust accumulation or sensor inaccuracies of up to 30 °C can compromise system stability, the proposed framework maintains operational integrity by limiting useful exergy destruction to a manageable 9.3%.