Abstract: Flow boiling is widely encountered in many industrial fields such as the steam generator in nuclear reactors. Nowadays, a lot of studies are focused on the bubble dynamics parameters due to the close relationship between the heat transfer and bubble behaviors. Bubble departure frequency is one of the important bubble dynamic parameters which reflects the bubble cycle in time and thus directly influences the bubble evaporation heat. In order to improve the predicting accuracy of bubble departure frequency and thus improve the accuracy of heat transfer calculation under flow boiling, the affecting factors of bubble departure frequency firstly through 13 groups of flow boiling independent experiments of water (506 experimental data points) were analyzed in this paper. It is found that bubble departure frequency decreases with the increasing of liquid subcooling, while it increases with the increasing of channel size, mass flux, wall superheating and pressure in which the effect of channel size is often ignored by many researchers. Then the accuracy of five existing prediction models was evaluated. The results show that none of them have high prediction accuracy in all experimental data points. Cole model and Zuber model both perform badly and have the absolute errors of 163.2% and 218.2% respectively and the relative errors of 134.2% and 205.1% respectively, because they only consider the effect of physical properties and thus they are more suitable for pool boiling. Brooks model and Basu model both overestimate the results with the absolute errors of 280.2% and 102.1% respectively and the relative errors of 235.2% and 37.1% respectively. The reason lies in that these two models are developed through their own experiment data and thus have limited application ranges. Chen model performs relatively better with the absolute error of 64.8%, but it largely underestimates the results with the relative error of -52.3% for the reason that this model is developed based on the pool boiling experiments of methane and ignored the effects of mass flux and liquid subcooling. Besides, the results show that most predicting models perform badly when the channel size is small which confirm the significant influence of channel size. Therefore, in this paper, a new prediction model was developed with higher predicting accuracy through dimensionless analysis which consideres the effects of channel size, physical properties, wall superheating and mass flux. This new model has the mean absolute error of 39.2% and mean relative error of -14.3% with the application ranges of bubble Reynold number 282 303, Bond number 1.9272 and Jakob number 0.0040.049.