联合时序分解和深度学习的高土石坝变形预测北大核心CSCD

针对高土石坝变形监测时间序列复杂的非线性、非平稳性等特点,提出了一种沉降变形预测的组合方法,可更好兼顾土石坝变形的长期发展趋势和波动特性。利用基于局部加权回归的季节性趋势分解法将变形监测历史数据分为趋势、周期和残差分量;采用长短期记忆神经网络模型分别学习趋势、周期和残差序列趋势特征并预测,汇总各分量预测结果得到大坝位移的预测值。为定量评价和比较预测结果,引入三个评价指标,并将预测结果与季节性差分自回归滑动平均模型、长短时记忆神经网络模型及其组合模型的预测结果进行对比分析。本文联合时序分解和深度学习的组合模型具有更高的预测精度和较好的稳定性,能够较好体现土石坝变形的长期趋势和随水位的波动特性。

Deformation prediction of high embankment dams by combining time series decomposition and deep learning

This paper develops a novel combined method of deformation forecasting for high embankment dam, considering the complex nonlinearity and non-stationarity of the time series, to improve the simulation of the long-term development trend and its fluctuation characteristics. Seasonal-trend decomposition based on the Loess smoothing is adopted to decompose the dam displacement time series into trend, seasonal, and remainder components. A long-short term memory neural network model is used to predict separately the three components that are then summed up to generate a displacement prediction. To evaluate and compare the prediction results quantitatively, three evaluation indicators are introduced and the results are compared with those of the SARIMA model, the LSTM neural network model, and a combined model of SARIMA-LSTM. New deformation model shows high accuracy and stability in the prediction. It is applicable to deformation with long-term trend and fluctuation characteristic with water level.