基于时间序列和CNN-LSTM的滑坡位移动态预测

Dynamic prediction of landslide displacement based on time series and CNN-LSTM

针对滑坡演化的动态特性以及静态预测模型的不足,提出一种基于时间序列与卷积长短期记忆神经网络(CNN-LSTM)的滑坡位移动态预测模型。首先,在利用曼-肯德尔(Mann-Kendall)趋势检验法检验滑坡位移的趋势特征的基础上,采用二次移动平均法将滑坡累积位移分解为趋势项位移和周期项位移,并采用最小二乘多项式函数拟合趋势项位移。然后,采用频谱图检验滑坡位移的周期特征,对周期项位移采用CNN-LSTM混合网络模型进行预测。最后,将两者预测值叠加得到累积位移预测值。该模型在中国三峡库区八字门滑坡得到了验证。实验结果表明,与当前常见的预测模型相比,所提方法的能有效对滑坡位移进行预测,且在周期项位移预测方面相较于经典BP神经网络,LSTM网络以及“门循环”(gated recurrent unit,GRU)网络模型,CNN-LSTM网络模型有更高的预测精度。

Aiming at the dynamic characteristics of landslide evolution and the shortcomings of static prediction model,a dynamic prediction model of landslide displacement based on time series and convolution long and short-term memory neural network(CNN-LSTM)is proposed in this paper.Firstly,based on the trend characteristics of landslide displacement tested by Mann-Kendall trend test method,the cumulative displacement of landslide is decomposed into trend term displacement and periodic term displacement by quadratic moving average method,and the least square polynomial function is used to fit the trend term displacement.Then,the periodic characteristics of landslide displacement are tested by spectrum diagram,and CNN-LSTM hybrid network model is used to predict the displacement of periodic term.Finally,the two predicted values are superimposed to obtain the cumulative displacement predicted value.The model has been verified by Bazimen Landslide in the Three Gorges Reservoir area of China.The experimental results show that compared with the current common prediction models,the proposed method can effectively predict the landslide displacement,and compared with BP neural network,LSTM network and gated recurrent unit(GRU)network model,CNN-LSTM network model has higher prediction accuracy in periodic term displacement prediction.