小样本下基于Transformer与树模型的临界热流密度预测及误差行为分析OA

To address the uneven distribution,inconsistent physical dimensions,and small-sample sensitivity of critical heat flux(CHF)experimental data under actual reactor operating conditions,this study establishes a research workflow integrating unified preprocessing,Transformer-based prediction,tree-model benchmarking,and error mechanism analysis.Based on a curated comprehensive CHF database,tube inner diameter,heated length,system pressure,mass flux,and outlet quality are selected as input features.Pressure units are unified,and samples with out-of-range outlet quality are marked for tracking.The results show that when the sample size is N<150,tree-based models,including Categorical Boosting,eXtreme Gradient Boosting,and Random Forest,exhibit more stable perfor-mance in terms of root mean square error and mean absolute error.When N increases to approximately 200-300,the Transformer model rapidly approaches and partially surpasses the tree-based models,demonstrating its potential for capturing high-dimensional nonlinear relationships.Residual analysis indicates that pressure and outlet quality are the dominant factors affecting prediction errors,while high-pressure and near-dryout conditions remain challenging for data-driven CHF prediction.By explicitly characterizing the small-sample performance curves,this study provides an alternative pathway for advancing data-driven CHF modeling toward engineering applications.