Abstract: A prediction model of final sulfur content is developed for hot metal pre-desulfurization by use of an improved BP neural network, especially for the desulfurization process by CaO+Mg powder co-injection in Benxi Steel Co. Ltd. To overcome the disadvantages of overmuch iterative repetition and slow convergence rate of normal BP algorithm, an approach to readjust adaptively the self-learning rate with self-learning for maximum error is used to improve the normal BP algorithm during modeling. The data from 1900 heats are used to train the model with other 100 heats randomly picked out as test samples. Test results showed that 12% of the predicted values got from the 100 samples are the same to the actually measured values, 89% have the error within 0.003% and the average error is 0.0020%.