Landslide displacement prediction model based on improved STL and multi-scenario verification

Landslide displacement exhibits dynamic, nonlinear and non-stationary characteristics. The traditional STL, Seasonal-Trend decomposition using LOESS (locally estimated scatterplot smoothing) only extracts periodic components from time series and ignores external driving factors such as rainfall and reservoir water level, resulting in insufficient decomposition accuracy. This paper proposes an improved STL (ISTL). By introducing external environmental factors and constructing a ridge regression model to replace the LOESS smoothing process that only relies on time series in traditional STL, it can better characterize the relationship between external factors and periodic displacement, and improve the accuracy and physical rationality of periodic component extraction. The model decomposes displacement into trend, periodic and random components. The trend displacement is predicted by the Gated Recurrent Unit (GRU), and the periodic displacement is predicted by a Multilayer Bidirectional GRU (MBi-GRU). Experimental results based on the ZG111 monitoring point of the Bazimen landslide in Hubei Province show that under the condition that the amplitude of the decomposed periodic displacement is about 20 mm, the Mean Absolute Percentage Error (MAPE) of the model for periodic displacement prediction is 2.438%, the Root Mean Square Error (RMSE) is 0.771 mm, and the coefficient of determination (R²) is 0.977. The model also performs well in trend displacement and total displacement prediction. Through generalization verification at the ZG110 monitoring point and other landslide points, the MAPE of its periodic displacement is all less than 0.826%, showing excellent generalization ability and practical value. Future work will integrate static geological factors (e.g., rock-soil shear strength, lithology, and topographic slope) with existing dynamic environmental factors to build a coupled dynamic-static multi-factor prediction framework, further improving the model's prediction accuracy and long-term stability.