A machine learning and CFD based approach for fouling rapid prediction in shell-and-tube heat exchanger

Abstract

Rapid prediction of fouling and heat transfer performance over the full life cycle of an industrial heat exchanger is critical. Establishing a rapid-prediction model of the heat exchanger to realize full life cycle monitoring is an effective approach to assistant its operation, decontamination, and maintenance. With the advancement of machine learning, rapid prediction of heat exchanger fouling level used deep learning methods has become a research hotspot. Based on machine learning and CFD methods, a rapid fouling prediction surrogate model for the industrial heat exchanger was proposed on the example of steam generators in nuclear power plants. The mathematical model of fouling layer thermal resistance was developed. The 3D numerical simulation under different fouling levels was carried out. The high fidelity CFD simulation database was built, and four deep learning models (BPNN, PSO-BPNN, CNN, RBFNN) were adopted. The fouling thermal resistance of SG could be predicted rapidly according to operating parameters. The root-mean-square error of the four neural networks is less than 10⁻⁷ K/W. BPNN with PSO algorithm achieves the best balance between calculation time and prediction accuracy. The anti-noise performance of the prediction surrogate model was evaluated at different noise level of actual operating parameters. When the noise level is 5 %, predicted R² remains at 0.7615 and the mean relative error is still less than 15 %. The low-cost and fast prediction surrogate model developed in this paper can provide an effective reference for the maintenance and decontamination of industrial heat exchanger.