基于MHA-BiLSTM的尾矿坝位移预测  

Study on Prediction of Tailings Dam Deformation Based on MHA-BiLSTM

作　　者：杨玉好 杨斌 胡军[1] 李铭 张壮超 YANG Yuhao;YANG Bin;HU Jun;LI Ming;ZHANG Zhuangchao(School of Civil Engineering,University of Science and Technology Liaoning,Anshan 114051,China)

机构地区：[1]辽宁科技大学土木工程学院,辽宁鞍山114051

出　　处：《有色金属工程》2024年第10期147-157,共11页Nonferrous Metals Engineering

基　　金：辽宁省教育厅面上项目(LJKZ0322);辽宁科技大学校青年项目(2020QN10);2024年度辽宁省高校基本科研业务费项目。

摘　　要：尾矿坝变形受多因素影响,针对传统预测方法受到数据复杂性和非线性关系的限制,导致预测精度不足的问题,提出多头注意力机制(Multi-Head Attention)和双向长短时记忆网络(BiLSTM)结合预测尾矿坝位移的方法。在预测中,首先利用Z-score和Savitzky-Golay滤波对原始数据消除异常值和噪声的干扰;然后,利用灰色关联度方法确定坝体位移影响因素;最后,采用MHA-BiLSTM模型对坝体位移进行预测。以辽宁省某尾矿库实测数据为例,为评估新模型的性能与传统BiLSTM模型进行对比,结果表明该方法能够更准确地预测出坝体位移变化情况。Tailings dam deformation is influenced by multiple factors.In response to the limitations of traditional prediction methods due to data complexity and non-linear relationships,which result in inadequate prediction accuracy,a method combining Multi-Head Attention mechanism and Bidirectional Long Short-Term Memory(BiLSTM)is proposed for predicting tailings dam displacement.In the prediction process,the original data is first processed using Z-score and Savitzky-Golay filtering techniques to eliminate disturbances caused by outliers and noise.Subsequently,the Grey Relational Analysis method is utilized to determine the factors influencing dam displacement.Finally,the MHA-BiLSTM model is employed to predict the dam displacement.Taking the measured data of a tailings pond in Liaoning Province as an example,the performance of the proposed model is compared with the traditional BiLSTM model.The results show that this method can predict the displacement of dam more accurately.

关 键 词：多头注意力机制 BiLSTM网络 Z-score去除异常值 Savitzky-Golay滤波 坝体位移预测 

分 类 号：TD926.4[矿业工程—选矿] TV649[水利工程—水利水电工程]