Application of Long Short-Term Memory Neural Network and Prophet Algorithm in Slope Displacement Prediction

Abstract

The slope displacement prediction is crucial for the development of an early warning system, which can help prevent or reduce losses of lives, properties, and the local environment. This problem is particularly important in the Three Gorges Reservoir (TGR) area, where the influence of geological, weather, and hydraulic conditions on landslides is significant. It is generally acknowledged that reservoir landslides are complex nonlinear systems with dynamic and inter-related features. However, most studies focus on how to express the static relationships between triggering factors and the landside displacement. In this paper, a long short-term memory (LSTM) neural network model was applied for predicting the total displacement of the Bazimen landslide, based on the decomposition of displacement time series. The accumulated displacement can be divided into two main parts: the trend and the periodic terms. The long-term trend was fit with a cubic nonlinear regression model; the residual one (the periodic displacement) was predicted via the LSTM model. By analyzing historical information and the Pearson correlation coefficient, a dynamic model was developed using six controlling factors. The good consistency between the predicted and monitored data proves the superiority of the model in predicting dynamic time-series problems. Compared with conventional static methods (i.e., MARS and SVM), the LSTM model can make full use of historical information due to its special “memory” structure. However, all these three methods can only perform well in forecasting one-step problems. To meet the requirement of multi-step forecasting, the Facebook Prophet model was also used in this study to predict landslide displacements with a longer period. The predicted results demonstrate the model’s superiority in efficiency and practice, at a cost of prediction accuracy.