关键词: 动态残差修正, 长短时记忆(LSTM)算法, 突发型滑坡, 位移预测, 流转训练, 经验模态分解(EMD)

Abstract:

In view of the sudden change of the displacement trend of the sudden landslide and the difficulty of the traditional LSTM method in the accuracy of displacement prediction, this paper proposes a displacement prediction method of the sudden landslide based on the dynamic residual correction LSTM network LSTM algorithm. First, the dynamic flow training was used to decompose the deformation rate obtained from the accumulated displacement into the periodic term and the trend term through EMD. Secondly, the trend term was predicted by polynomial and the period term was predicted by dynamic LSTM, and the main predicted deformation rate was obtained from the sum of the two terms, Subsequently, The residual term was obtained by comparing the measured rate with the main predicted deformation rate, and the residual LSTM network of "dynamic flow training" was established to predict the residual rate. Finally, taking a sudden landslide as an example, the method proposed in this paper was used to predict the displacement of the landslide. The results show that the MAE, MAPE, RMSE, and R² indicators based on the dynamic residual correction LSTM algorithm are 43.843%, 1.901%, 79.394%, and 0.960%, respectively, which are higher than traditional LSTM prediction methods.

Key words: dynamic residual correction, long short term memory (LSTM) algorithm, sudden landslide, displacement prediction, flow training, empirical mode decomposition (EMD)