The critical heat flux (CHF) for subcooled water flow boiling is experimental studied under high heat fluxes conditions (10∼15 MW/m²), and the prediction methods are focused on. The effects of thermodynamic vapor quality, mass flux and pressure on CHF of subcooled boiling are analyzed. It is found that an increase in thermodynamic quality decreases the CHF linearly. The CHF increases with the rise of mass flux; however, when thermodynamic quality approaches the saturation point, the trend gradually decreases. The subcooled boiling CHF data in open literatures are collected to form a dataset of subcooled boiling of high heat flux with the paper data, including 709 data sets. And the empirical correlations and neural network models are adopted to make prediction research, and the prediction performance of seven empirical correlations and three neural networks model (BP neural network, GA-BP neural network, MEA-BP neural network) were quantitively evaluated. The results indicate that the prediction performance of machine learning has improved significantly compared to traditional correlation, and the MEA-BP neural network has the best prediction performance, the Mean Absolute Error is 15.61% and the Root Mean Square Error is 21.56%。