A step-type landslide displacement prediction model based on creep trend influence and feature optimization algorithm

In the construction process of the landslide displacement prediction model，in order to clearly separate the landslide displacement parts affected by various features to enhance the mechanical interpretability and generalization ability of the displacement prediction model，a step-type landslide displacement prediction model based on the influence of creep trend and feature optimization algorithm is proposed. Firstly，the Particle Swarm Optimization-Based Variational Modality Decomposition(PSO-VMD) Method is used to decompose the training data into other data curves and trend displacement with influence of Nishihara creep model. Secondly，the external influence features of landslide displacement are comprehensively selected，other data curves are separated and reconstructed with the genetic feature optimization algorithm(GA) to obtain the fluctuation displacement with Fourier characteristics and the influence displacement that maximizes the correlation with external influence features. The Nishihara creep Taylor expansion polynomial model，Fourier function model and Convolutional Neural Network-Squeeze Excitation Attention Module-Gated Recurrent Unit Model(CNN-SE-GRU) are used to respectively fit and predict the trend displacement，fluctuation displacement and influence displacement. The errors of each displacement model training fit are defined as the random displacement；it is predicted by the Gray Wolf Optimization-Based Extreme Learning Machines model under the constraint of kernel density estimation(KDE-GWO-ELM). Finally，the prediction of step-type landslide displacement is accomplished by superimposing all displacement prediction model results. Taking the step-type landslide- namely Baishuihe landslide as an example，the cumulative horizontal displacement data of monitoring points ZG118 and ZG93 from November 2005 to October 2009 are selected as training data，and the data from November 2009 to July 2010 are selected as prediction data for research. The prediction results show that the root mean square error(<italic>RMSE</italic>) of monitoring points ZG118 and ZG93 under the prediction model 5.85 mm and 10.61mm respectively，the mean absolute percentage error(<italic>MAPE</italic>) are 0.27% and 0.43% respectively. Compared with the CNN-SE-GRU/LSTM model，GWO-ELM model，VMD-GWO-ELM model and EEMD-CNN-GRU model，the novel prediction model not only improves the prediction accuracy，but also shows clearly the influence of time features，external influence features and error features for prediction model. Thus，it has excellent interpretability and generalization ability.