Modular surrogate modeling-based optimization framework for thermohydraulic systems assisted by machine learning

Improving performances of thermohydraulic systems (THSs) is crucial for energy saving. Tradeoffs between model accuracy and optimization complexity are required for system optimization within acceptable costs, leading to less-reliable results in practice. This paper introduces a modular surrogate modeling-based optimization framework for THSs to tackle this challenge. THSs are first decomposed into components, whose surrogate models are constructed based on datasets collected from experiments or simulations. Component models are integrated to build the system model, which is finally used for system optimization. The optimization complexity is effectively reduced with the accuracy kept. Two cases are investigated for validation. A heat exchanger network with supercritical carbon dioxide is first considered with datasets generated from simulation. The obtained optimal design consumes 27.5% less power compared with that of the reference design. The robustness of optimization via the framework also outperforms the physics-based model substantially. An integrated cooling system with two operation modes is considered next with datasets collected from experiments. The optimal system operation guided by the system model reduces power consumptions by 47.3 and 36.4% under two modes. Finally, other advantages of the framework such as dataset scale, reliability, interpretability, and flexibility are elaborated to highlight its superiority.