用变化较大的、精度不高的数据组建立多层反向传播神经网络(BPANN)模型,用以模拟印度纳尔默达河两个小流域Banjar~Hridaynagar和Narmada~Manot的3种时间尺度(周、10d、月)的降雨径流过程。采用梯度下降优化技术建立BPANN径流模型,并通过交叉检验进行概化。用变化相对较大、精度不高的数据资料建立起来的BPANN,在几乎所有情况下,计算迭代次数较少,概化程度较高。计算结果表明,BPANN模型的性能优于线性转换函数模型(LTF)。 A multi-layer backpropagation neural network (BPANN) model was established with large and poorly-defined datasets to simulate three time scales of Banjar ~ Hridaynagar and Narmada ~ Manot in two small catchments of the Narmada River in India (Week, 10d, month) rainfall runoff process. BPANN runoff model was established by using gradient descent optimization technology and generalized by cross-test. In almost all cases, BPANN, which is established by using relatively large and poorly-defined data, has fewer computational iterations and higher degree of generalization. The calculation results show that the performance of BPANN model is better than the linear transfer function model (LTF).