Prediction of landslide displacement with step-like curve using variational mode decomposition and periodic neural network

Landslide deformation characterized with step-like curves often presents periodicity implicitly. This paper proposed a novel data-driven approach that adopted periodic neural network (PNN) and variational mode decomposition (VMD) to conduct displacement prediction based on the intrinsic seasonality of step-like landslide displacement. PNN was a novel neural network designed for capturing the seasonality of the time series. Firstly, the initial displacement would be decomposed into trend component, periodic component, and random component using the variational mode decomposition (VMD). Then, the external triggering factors were also decomposed by VMD into several subsequences. Subsequences with periodic and random characteristics were selected as the input datasets to forecast the periodic and random components by PNN. Finally, the total displacement was obtained by superimposing the three predictive components to validate the model performance. The Baishuihe landslide was taken as a case study to validate the high effectiveness and efficiency of our method. The result proved that our new model presented satisfactory prediction accuracy without complex training process. Meanwhile, PNN performed a strong robustness to the missing values due to the advantage of its structure. In addition, we clarified a corrective data processing mode as "strict" mode: the dataset has to be divided into training and validation sets firstly to avoid the leakage of the future data.