Interval Estimation of Landslide Displacement Prediction Based on Time Series Decomposition and Long Short-Term Memory Network

Abstract

Interval estimation of landslide displacement prediction is significant for landslide early warning. The goal of this paper is to improve the accuracy of landslide displacement point prediction and quantify the uncertainties associated with the predicted values. To do so, a coupling prediction model based on double moving average (DMA) method and long short-term memory (LSTM) network is investigated. The DMA method is employed to decompose cumulative displacement of landslide into trend and periodic displacements, while the LSTM network is adopted to model and predict these two sub dis-placements. The sum of predicted sub displacements is considered as predicted cumulative displacement. Further, the probability estimation theory is utilized to derive confidence intervals that quantify the uncertainties of the point prediction. The proposed approach was validated on Baishuihe landslide in Three Gorges Reservoir area of China. Results show that the LSTM network performs better than support vector machine and Elman network, while the DMA decomposition method outperforms single moving average method. As a consequence, the coupling prediction model of DMA and LSTM network is a better solution for the point prediction of landslide displacement. Furthermore, the proposed probability estimation method can construct high-quality confidence intervals.