基于小波优化LSTM-ARMA模型的岩土工程非线性时间序列预测

Prediction for Nonlinear Time Series of Geotechnical Engineering Based on Wavelet-Optimized LSTM-ARMA Model

摘要:

为了更精确地预测岩土工程应力、变形等的非线性时间序列，提出了基于小波优化的长短时记忆神经网络?自回归滑动平均模型（LSTM-ARMA）预测模型。首先使用小波分析将监测序列分解成趋势项和噪声项，采用LSTM网络滚动预测趋势项、ARMA模型预测噪声项，并将趋势项预测值与噪声项预测值之和作为总的时间序列预测值。在此基础上，以上海云岭超深基坑工程为案例进行了基坑地表沉降分析，结果表明组合模型的预测精度要高于单一LSTM模型且更加稳定；进一步采用弹塑性有限元对基坑开挖诱发的地表沉降进行了预测，并与人工智能预测结果进行对比，验证了人工智预测模型的合理性。分析表明，当后续工况与前置工况所诱发的变形机理突变时，人工智能预测误差增大，但伴随后续工况的推进，人工智能预测误差将逐渐减小。

Abstract:

In order to predict the nonlinear time series of geotechnical engineering more precisely， a wavelet-optimized LSTM-ARMA model is proposed. First， the monitoring series are decomposed into a trend term and a noise term through wavelet analysis. Then， the trend term is predicted by the long short-term memory network （LSTM）， while the noise term by the autoregressive moving average model （ARMA）. Finally， the sum of the predicted values of both terms is taken as the total predicted results. The performance of the method is validated through the case analysis of an ultra-deep foundation pit which also indicates that the combined model gives a more precise and stable prediction than the LSTM network. Besides， the elastic-plastic finite element method is also used to predict the ground settlement induced by foundation pit excavation， and its results are compared with those of the artificial intelligence method， verifying the rationality of the latter. The analysis shows that the prediction error of the artificial intelligent method will increase significantly when the deformation mechanisms of the previous and the subsequent working conditions change suddenly， but it will decrease gradually with the progress of the subsequent working conditions.