Abstract: In order to improve the hit rate and accuracy of temperature and composition of molten steel during RH refining process, the prediction model for the end point temperature and composition of molten steel was established based on integrating of metallurgical mechanism and NARX Neural network. The prediction model program that was able to calculate the amount of oxygen and alloy and to predict the end point was developed by utilizing of Visual Basic 6.0 and Matlab neural network toolbox. The model showed high prediction accuracy with over 85 % hit rate of temperature and composition at the same time. The rate of temperature errors less than 5 ℃ reached 90 % and all the carbon mass fraction prediction errors were within 5×10^-6. The rates of Si，Mn，P and Al_s prediction errors less than ±5 % exceeded 90 %. In addition, the rates of oxygen, lowcarbon ferrosilicon, iron phosphate, aluminum and microcarbon ferromanganese amount prediction error within －3 % to 7 % were 90 %，75 %，75 %，95 % and 70 %, respectively.

Key words: RH refining, prediction model, temperature, composition, alloying, neural network