A universal ANN-based approach predicting PCHEs’ off-design performance across various operating conditions of sCO2 RCBCs

Abstract

This paper proposes a universal approach that employs artificial neural networks (ANN) to predict the off-design performance of printed circuit heat exchangers (PCHEs) using supercritical carbon dioxide (sCO2) fluid. PCHEs are widely studied precoolers and recuperators of sCO2 power cycles, for which predicting their off-design heat transfer performance is challenging due to the fluid’s dramatic non-linear property variations. Due to their narrow valid range and constrained expressing ability, the existing empirical correlations might introduce significant out-of-scope errors that result in diverged simulation results, slow convergence and abnormal quitting. This work addresses the issue by proposing an ANN-based Nusselt number correlation, ANNu, to predict the off-design performance of PCHEs quickly and universally over the vast operating range of sCO2 cycles. Multiple techniques were adopted to promote ANNu’s prediction accuracy and generalization, such as consolidating 494 training data, data-driven input selection, and global optimization for the networks’ architecture and hyperparameters. This paper comprehensively evaluates ANNu’s performance by comparing its predictions with ten empirical correlations over four typical real-world scenarios, where ANNu features an overall MSE of 83.4222 and overall MAPE of 16.285 % that rank the second best and third best among eleven candidates (even overpassing two references), respectively. Moreover, ANNu exhibits good generalization and reliability, featuring MAPEs ranging from 9.89 % to 27.73 % when processing unmet experiment data from sCO2 PCHE applications.