基于改进知识蒸馏的干涸后传热预测方法OA

Objective To address the issues of limited experimental data and insufficient cross-domain generalization in modeling post-dryout heat transfer coefficients,an improved knowledge distillation-based continual learning method is de-veloped.Method Based on two independent post-dryout heat transfer experimental datasets,the proposed approach first trains a model using the dataset from Institution Ⅰ and employs a Conditional Tabular Generative Adversarial Network(CTGAN)to generate synthetic feature samples for data augmentation and automatic labeling.Then,on the basis of the Institution Ⅰ model,a model for Institution Ⅱ is trained through joint optimization of logit distillation and feature distilla-tion,enabling collaborative retention and adaptive transfer of old and new knowledge.Results The proposed model achieves a mean absolute percentage error(MAPE)of 1.92%and a coefficient of determination(R2)of 0.999 6 on an in-dependent test set.Compared with fine-tuning and feature extraction methods,it demonstrates superior prediction accuracy and generalization capability.Conclusion The proposed method effectively enhances the predictive performance of post-dryout heat transfer models while ensuring data security,providing a feasible technical route for intelligent heat transfer modeling under complex conditions in nuclear energy systems.