论文部分内容阅读
在对某基坑工程采用BP神经网络模型预测基坑开挖引起地表变形的分析中,考虑到现有模型可能会遇到预测结果跳不出训练样本以及训练时间较长的问题,提出采用Matlab中的mapminmax函数进行归一化处理,并基于牛顿法、共轭梯度法和L-M法三种数值优化方法对BP网络训练算法进行了改进。研究结果表明:与常用的基于梯度下降原则相比,改进后的BP神经网络在训练时间和预测误差方面均有明显的优势,采用L-M法的神经网络在训练样本时的迭代次数最少为74次,采用共轭梯度法的预测结果与实测结果的误差最大为2.4%,而采用牛顿法神经网络的预测值则比较均衡,预测结果相对最佳。
In a foundation pit project using BP neural network model to predict excavation caused surface deformation analysis, taking into account the existing model may encounter the prediction results can not jump out of training samples and training time is longer, the proposed use of Matlab The normalized mapminmax function is used to improve BP neural network training algorithm based on three numerical optimization methods: Newton method, conjugate gradient method and LM method. The results show that compared with the commonly used principle of gradient descent, the improved BP neural network has obvious advantages in terms of training time and prediction error. The neural network with LM method has at least 74 iterations in training samples , The error between the predicted result and the measured result using conjugate gradient method is 2.4%, while the predicted value using Newton method neural network is more balanced, and the prediction result is the best.