针对神经网络建模效果对网络结构、训练方法过于敏感的缺陷,提出一种结构自适应神经网络模型及其训练方法.模型具有双网结构并以“提前终止法”训练,一定程度上降低了建模效果对网络结构的敏感性;模型结构根据建模数据的噪声方差、模型当前误差等信息自适应调整,进一步提高了模型的建模效果,同时具有较高的时间效率.仿真结果表明,该方法弥补了提前终止等传统方法的部分不足,具有较好的效果. Aiming at the defect that neural network modeling effect is too sensitive to the network structure and training method, an adaptive neural network model and its training method are proposed. The model has a dual-network structure and is trained by “early termination method” to a certain extent Which reduces the sensitivity of the modeling effect to the network structure. The model structure adaptively adjusts according to the noise variance of the modeling data and the current error of the model to further improve the modeling effect of the model, and has high time efficiency. It shows that the method can make up for some shortcomings of traditional methods such as early termination and has good effect.