In order to reduce the iron loss and improve the dephosphorization efficiency of the converter for dephosphorization by the full triple stripping process, a model, based on the oxygen balance mechanism, is bulit to predict the end point FeO content and the Levenberg-Marquardt neural network algorithm is adopted in this model. The calculation of the oxide mass (FeO, CaO, SiO₂, MgO, MnO, P₂O₅, Al₂O₃) with the oxide balance mechanism model and the tapping temperature are used as inputs to the neural network toolbox to train the network with minimum error. The results show that the heat with relative error of 10% between the predicted value and the measured value of FeO is up to 85%.This proves that the FeO prediction hit rate of the model is high, and can provide theoretical basis for production on site.