样本构成、网络结构、学习算法是影响前馈网络应用的三大关键问题。本文提出了一种基于门限自回归ＴＡＲ模型的非线性时间序列预测前馈网络方法，该方法利用ＴＡＲ模型的门限值对训练样本进行分群，依据ＴＡＲ模型的阶数、训练样本数等确定前馈网络结构，网络学习算法采用基于梯度法和共轭梯度法相结合的联合梯度算法。应用研究表明该方法有效地改善了网络的泛化性能，提高了预报精度，同时也缩短了网络的训练时间。 Sample structure, network structure and learning algorithm are the three key issues that affect the application of feedforward network. This paper presents a nonlinear time-series prediction feed-forward network based on threshold auto-regressive TAR model, which uses TAR threshold to classify training samples. According to the order of TAR model and the number of training samples, The feed-in network structure and the network learning algorithm use a joint gradient algorithm based on the gradient method and the conjugate gradient method. Applied research shows that this method can effectively improve the generalization performance of the network, improve the forecasting accuracy and shorten the training time of the network.