改进 LSTM 沉降预测模型在既有线铁路结构沉降预测中的应用

作者：严申华¹,张家奇¹,张拯²,褚韶波¹,钱华¹,陈愿¹

作者简介：

严申华，工程师，E-mail: 2822174313@ qq. com

作者单位：

1.中国铁路上海局集团有限公司杭州铁路枢纽工程建设指挥部，浙江杭州 310000;2.成都西南交大技术转移中心有限公司，四川成都 610036

基金项目：

∗中国铁路上海局集团有限公司科技项目：铁路高架站房改造多源扰动下既有线铁路设施设备及钢混基础结构智能监测预警研究（E2024120）

摘要：

以沪昆铁路义乌高架站房建设工程杭长道工程为例，构建了新型耦合残差连接与多头注意力机制 LSTM,即既有线铁路结构变形预测网络（Res-MHA-LSTM），实现了准确地既有线铁路结构沉降预测。对比不同测点下的5 种常见时序预测网络的预测结果趋势和取值，并通过多种指标对结果进行了量化分析，验证了 Res-MHA-LSTM网络在多个测点数据下的预测准确性和可行性。在 HC-2 测点数据下，Res-MHA-LSTM 方法在模型预测结果的平均绝对误差、均方根误差和平均绝对百分比误差分别是 0.042,0.079,0.45,相比 LSTM 预测模型，分别降低了23.09%,23.51%和 87.69%.决定系数为 0.91,提升了 7.4%,相对其他试验模型也有提升。

关键词：铁路；沉降；变形；预测

English：

Taking the Hangchang Road project of Yiwu Elevated Station building on Shanghai-KunmingRailway as an example, a new type of coupled residual connection and multi head attention mechanismLSTM, namely the Res-MHA-LSTM,was constructed to accurately predict the structural settlement ofexisting railway lines. Comparing the prediction results trends and values of five common time-seriesprediction networks under different measurement points, and quantitatively analyzing the results throughmultiple indicators, the accuracy and feasibility of the Res-MHA-LSTM network in predicting data frommultiple measurement points were verified. Under the HC-2 measurement point data, the averageabsolute error, root mean square error, and average absolute percentage error of the Res-MHA-LSTMmethod in the model prediction results were 0.042, 0.079 and 0.45, respectively. Compared with theLSTM prediction model, they decreased by 23.09%, 23.51% and 87.69%, respectively. Thecoefficient of determination is 0.91, an increase of 7.4%,which is also an improvement compared toother experimental models.

[1]	TUCHO A,INDRARATNA B,NGO T. Stress-deformation analysisof rail substructure under moving wheel load[ J]. Transportationgeotechnics,2022,36:100805.
[2]	VARANDAS J N,PAIXÃO A,FORTUNATO E,et al. Long-termdeformation of railway tracks considering train-track interactionand non-linear resilient behaviour of aggregates—a 3D FEMimplementation [ J ]. Computers and geotechnics, 2020,126:103712.
[3]	CHAWLA S, SHAHU J T, KUMAR S. Analysis of cyclicdeformation and post-cyclic strength of reinforced railway trackson soft subgrade [ J ]. Transportation geotechnics, 2021,28:100535.
[4]	CHEN Z W. Evaluation of longitudinal connected track undercombined action of running train and long-term bridgedeformation[J]. Journal of vibration and control,2020,26（7 / 8）：599-609.
[5]	ZAMBRANO O A,TRESSIA G,SOUZA R M. Failure analysis ofa crossing rail made of Hadfield steel after severe plasticdeformation induced bywheel-rail interaction [ J]. Engineeringfailure analysis,2020,115:104621.
[6]	XING Y, CHEN Y, HUANG S P, et al. Research on damdeformation prediction model based on optimized SVM [ J ].Processes,2022,10（9）：1842.
[7]	WAN P,ZOU H,WANG K L,et al. Research on hot deformationbehavior of Zr-4 alloy based on PSO-BP artificial neural network[J]. Journal of alloys and compounds,2020,826:154047.
[8]	ZHOU Z W,GONG H Y,YOU J,et al. Research on compressiondeformation behavior of aging AA6082 aluminum alloy based onstrain compensation constitutive equation and PSO-BP networkmodel[J]. Materials today communications,2021,28:102507.
[9]	ZHANG Y G, TANG J, LIAO R P, et al. Application of anenhanced BP neural network model with water cycle algorithm onlandslide prediction [ J]. Stochastic environmental research andrisk assessment,2021,35（6）：1273-1291.
[10]	LIN C,WENG K L,LIN Y L,et al. Time series prediction of damdeformation using a hybrid STL-CNN-GRU model based onsparrow search algorithm optimization [ J ]. Applied sciences,2022,12（23）：11951.
[11]	夏超鹏，洪宇，刘新山，等。基于 LSTM 深度学习网络的中长期电力负荷预测【J].自动化与仪器仪表，2024（7）：343-346.XIA C P,HONG Y,LIU X S,et al. Based on LSTM network formedium-term and long-term power prediction based on pytouch[J]. Automation & instrumentation,2024（7）：343-346.
[12]	谭茜元，谷佩奇，张悦悦。基于灰狼优化算法优化 LSTM 网络的短期负荷预测【 J].现代工业经济和信息化，2024,14（7）：267-268,271.TAN X Y,GU P Q,ZHANG Y Y. Optimising LSTM networks forshort-term load forecasting based on the greywolf optimisationalgorithm[J]. Modern industrial economy and informationization,2024,14（7）：267-268,271.
[13]	CHEN J,LI A Y,BAO C Y, et al. A deep learning forecastingmethod for frost heave deformation of high-speed railway subgrade[J]. Cold regions science and technology,2021,185:103265.
[14]	WANG X,BAI Y,LIU X. Prediction of railroad track geometrychange using a hybrid CNN-LSTM spatial-temporal model [ J].Advanced engineering informatics,2023,58:102235.
[15]	袁利伟，杨柳，何涛，等。结合 CNN-LSTM 优化模型在尾矿坝浸润线预测中的应用【 J].安全与环境学报，2024,24（ 10）：3770-3779.YUAN L W,YANG L,HE T, et al. Application of CNN-LSTMoptimization model in predicting tailing dam saturation lines[ J].Journal of safety and environment,2024,24（10）：3770-3779.
[16]	HAN K, XIAO A, WU E, et al. Transformer in transformer[J].Advances in neural information processing systems, 2021, 34:15908-15919.
[17]	WANG J, LI C, LI L, et al. InSAR time-series deformationforecasting surrounding Salt Lake using deep transformer models[J]. Science of the total environment,2023,858:159744.
[18]	周康敏，程康，曾少翔，等。基于深度残差 LSTM 的盾构姿态预测【J].隧道建设（中英文），2024,44（8）：1643-1651.ZHOU K M,CHENG K,ZENG S X,et al. Prediction of shieldattitude using deep residual long short-term memory model[ J].Tunnel construction,2024,44（8）：1643-1651.
[19]	LI J,WANG X, TU Z P, et al. On the diversity of multi-headattention[J]. Neurocomputing,2021,454:14-24.[20] YU Y, SI X S, HU C H, et al. A review of recurrent neuralnetworks: LSTM cells and network architectures [ J ]. Neuralcomputation,2019,31（7）：1235-1270.

岩土与地下工程 > 2026年 第卷 第02期