本文构造了一种新的基于线性模型、多层前向网络的混合结构神经网络模型,并提出了相应的非迭代快速学习算法。该学习算法能够根据拟合精度要求,运用线性最小二乘法确定相应的最佳网络权值和线性部分的参数,并自动确定最佳的隐层节点数。与BP网络的比较结果表明,本文提出的混合结构前向神经网络的快速学习算法无论在拟合精度、学习速度、泛化能力、还是隐节点数均显著好于BP算法。 In this paper, we construct a new hybrid neural network model based on linear model and multi-layer forward network, and propose a corresponding non-iterative fast learning algorithm. According to the requirement of fitting accuracy, the learning algorithm can determine the optimal network weights and linear part parameters by using linear least square method and automatically determine the optimal number of hidden nodes. Compared with BP network, the results show that the fast learning algorithm proposed in this paper is significantly better than BP algorithm in terms of fitting accuracy, learning speed, generalization ability and number of hidden nodes.