优化BP神经网络模型的边坡安全多因素滚动预报

Optimized BP neural network model for slope safety multi-factor rolling prediction

为提高BP神经网络模型在边坡安全预报中的普适性及预报精度,对原始数据进行卡尔曼滤波去噪处理,通过确定最优时间跨度大小、隐含层神经元个数,保障模型最优拓扑结构,提出基于BP神经网络模型的时间序列数据最优时间跨度边坡安全滚动预报方法。选取绥江县城新址边坡工程的地表沉降,结合水位、应力两种主要影响因素监测数据为预报指标,首先对监测数据采用Kalman滤波进行预处理,其次进行输入层时间跨度划分,求得三类数据最优BP神经网络拓扑结构,构建KF-BP神经网络预报模型进行适应性分析并预报。结果表明,①确定三类数据最佳时间跨度大小分别为5、8和7期,构建最优拓扑结构的BP模型预报精度均表现较好;②经卡尔曼滤波削弱原始监测数据中的干扰和噪声后,数据预报结果得到明显提高,其中对水位监测数据的预报效果显著,其平均绝对误差(MAE)、平均绝对百分比误差(MAPE)、均方根误差(RMSE)、相关系数(R)分别为0.009、0.01%、0.012和0.945;③采取3种类型监测数据对边坡安全进行多因素分析和预报,证明所构建模型对多类数据预报的普适性的同时,从整体上评价边坡安全,为边坡工程的安全预报及预警提供借鉴。

In order to improve the universality and accuracy of slope safety prediction by BP neural network model,the original data were de-noised by Kalman filter,and the optimal time span size and the number of hidden layer neurons were determined to ensure the optimal topology of the model.A rolling prediction method of slope safety with the optimal time span based on BP neural network model was proposed.This paper selected the surface settlement of the slope project at the new site of Suijiang County,combined with the monitoring data of the two main influencing factors,water level and stress,as the prediction index.Firstly,the monitoring data was preprocessed by Kalman filter,and then the input layer time span was divided to obtain the optimal BP neural network topology of the three types of data.The KF-BP neural network prediction model is constructed for adaptive analysis and prediction.The results show that,①The optimal time span of the three types of data is determined to be 5,8 and 7 epochs respectively,and the prediction accuracy of the BP model with the optimal topology structure is good;②After the interference and noise in the original monitoring data are weakened by Kalman filtering,the data prediction results are significantly improved,and the prediction effect of the water level monitoring data is significant.The mean absolute error(MAE),mean absolute percentage error(MAPE),root mean square error(RMSE)and correlation coefficient(R)were 0.009,0.01%,0.012 and 0.945,respectively;③Multi-factor analysis and prediction of slope safety with three types of monitoring data are adopted to prove the universality of the constructed model for the prediction of multiple types of data,while evaluating slope safety as a whole,providing reference for the safety prediction and early warning of slope engineering.