基于传递函数优化后BP神经网络对泥石流的危险性预测

泥石流危险性预测的可靠性是防治工程建设与减灾救灾相关工作部署的关键，基于Back Propagation神经网络的预测方法，是目前实现危险性等级划分的有效方法之一。利用BP神经网络算法的非线性逼近能力，挑选陇南白龙江小流域26条典型泥石流沟道，结合当地实际情况，选取泥石流危险性的8个主要因素为输入层神经元，以样本数据危险等级为输出神经元，在测试单、双层隐含层网络性能的基础上，提出9种工况组合的传递算法搭配方案，利用L-M算法搜索最优解或者近似最优解，总结传递算法对泥石流预测模型精度的影响及算法的选择顺序。实验结果显示，隐含层采用tansig函数，输出层采用logsig函数，其模型总体误差最小，模型的R训练集、R验证集较大与R测试集分别为0.983 61、0.709 17和0.960 52,准确率达到96.1%。由此可见，选择合适的传递函数可提高网络模型的精准度，能准确划分泥石流风险等级。

Hazard prediction of debris flow based on BP neural network optimized by transfer functio

Abstract] The reliability of landslide risk prediction is the key to the deployment of prevention and control project construction and disaster relief and mitigation, and the prediction method based on BP neural network is one of the effective methods to achieve the classification of risk levels. This paper uses the nonlinear approximation ability of BP neural network algorithm to select 27 typical mudslide channels in the small watershed of Bailongjiang River in Longnan Province, combines the actual local conditions, selects the 8 main factors of debris flow risk as input layer neurons, takes the sample data danger level as output neurons, and proposes a transfer function collocation scheme for 9 combinations of working conditions on the basis of testing the performance of single and double layer implicit layer networks, and uses L-M algorithm to search for the optimal solution or approximate optimal solution. Summarize the influence of the transfer function on the accuracy of the debris flow prediction model and the selection order of the functions. Experimental results show that the implicit layer uses the tansig function, the output layer uses the logsig function, the overall error of the model is the smallest, the R training set, R verification set and R test set of the model are 0.98361, 0.70917 and 0.96052, respectively, the accuracy rate reaches 96.1%, and the selection of the appropriate transfer function can improve the accuracy of the network model.