A guideline study for optimal machine learning approaches to predict the off-design performance of sCO2-PCHEs

As research increasingly focuses on predicting the off-design performance of PCHEs using machine learning (ML) approaches, there is a clear need to identify the most cost-effective method that offers optimal accuracy and generalization across a wide range of operating conditions in sCO₂ power cycles. In response, this study establishes highly accurate machine learning proxies, including XGBoost, SVMs, BPNNs, and FCDNs for comparison, aiming to rank and identify the one with the best accuracy and generalization for different operating conditions of PCHEs. First, multiple datasets concerning the geometric parameters and working conditions of PCHEs are consolidated. Two data-driven techniques are introduced to correlate both inputs and outputs, identifying the most pertinent parameter. Then, the sensitivity of ML models to hyperparameter, split ratio, and data distribution is evaluated. Finally, a mutual assessment is conducted to rank their accuracy and generalization when predicting the heat transfer coefficient across three real-world operating scenarios. This study concludes that: 1) BPNN and XGBoost emerge as the best performers with the highest accuracy, achieving a MAPE of 2.938 % and 3.074 %, respectively, among eleven approaches with four ML models and seven Nu correlation-based models included; 2) BPNN and XGBoost exhibit strong generalization ability, demonstrating accuracy losses of less than 0.057 and 0.244, respectively, against unseen data; 3) unlike FCDN, which requires substantial computational time, or SVM, which produces significant errors, BPNN and XGBoost are more reliable and efficient in forecasting sCO₂-PCHEs’ performance for future power cycle control studies that involve vast operating conditions and numerous rounds of optimization search.