Coke rate is a very important technique index in the processing of metallurgical, and it is also an important goal that should be reached and controlled in practice.The blast furnace is a countercurrent heat and mass exchange reactor involving the solid, liquid and gaseous phases. Using computer encoded mathematical and statistical methods can not get the precise result. An improved 9-9-1 BP(Back propagation) neural network was trained and used in the prediction of the coke rate. The result indicates that the BP nets can predict coke rate accurately and the error between prediction and real coke rate less than 2%. And the use of a hybrid model in actual on-line intelligence control was also discussed.