讨论了权值初始化、变量的预处理、学习过程参数的自适应调节、网络拓扑结构等因素对学习和推广的影响 ,提出了一种改进的BP神经网络学习算法 ,在很大程度上改善了学习效率。采用改进的带有 8个输入变量的BP神经网络算法和自适应残差补偿算法建立吹炼终点组合预报模型。利用某厂实际生产数据进行仿真运行的结果表明 ,利用该组合预报模型得到的平均相对预测误差为 1 .2 % ,最大误差为 4%。 The influence of initialisation of weights, pretreatment of variables, adaptive adjustment of learning process parameters and network topology on learning and promotion are discussed. An improved BP neural network learning algorithm is proposed, which greatly improves Learning efficiency. A modified BP neural network algorithm with eight input variables and an adaptive residual error compensation algorithm are used to establish the combined forecasting model of blowing end point. The simulation results of the actual production data of a factory show that the average relative prediction error obtained by using the combined forecasting model is 1.2% and the maximum error is 4%.