根据Pad近似理论,构造出一类前向有理式神经网络.该网络采用四层结构,其中第一层(输入层)和第三层采用线性激励函数,第二层采用幂激励函数,第四层(输出层)采用分数函数(或称除法函数)作为激励函数.依据梯度下降法思想,推导了其权值修正的迭代公式.针对迭代方法收敛速度慢、易陷入局部极小点等缺点,进一步推导出了基于伪逆的权值直接确定方法,该方法避免了冗长的迭代过程.仿真和预测结果均表明Pad有理式神经网络及其权值直接确定法具有较好的计算速度和更高的逼近与预测精度. According to the approximation theory of Pad, a class of forward rational neural networks is constructed. The network uses a four-layer structure in which the linear excitation function is used for the first layer (input layer) and the third layer, the power excitation function is used for the second layer, According to the idea of ​​gradient descent method, the iterative formula for weight correction is deduced. For the shortcoming of the iterative method, such as slow convergence speed and easily falling into local minimum point, the fourth layer (output layer) adopts the fraction function (or division function) , A direct method of weight determination based on pseudo-inverse is further deduced, which avoids lengthy iterative process.The simulation and prediction results show that the Pad-rational neural network and its direct weight determination method have better computational speed and Higher approximation and prediction accuracy.