Landslide Displacement Prediction Based On Hybrid Machine Learning

Landslides are the most frequent,common,destructive and extensive geological disasters.Every year in China,landslides involve more than 20,000 casualties and seriously threaten the safety of people’s lives and property.The Three Gorges area has become a typical area studied by many scholars.The effectiveness of landslide prediction depends on selection of the trigger factors and the construction of high-performance prediction model.In the traditional prediction model,the sensitivity of influencing factors to the model is not considered,and the key trigger factors cannot be screened out.In addition,the existing prediction models lack of fine-grained analysis of features and limitations of models,which make it difficult to effectively predict landslides.Based on the above problems,the main research is as follows:(1)In order to solve the problem that the traditional methods did not consider the influence of the influencing factors on sensitivity of the landslide displacement and random fluctuation occurred during landslide creep,a prediction method of grey extreme learning machine based on double exponential smoothing was proposed.Firstly,the trend term and periodic term of cumulative displacement are extracted by double exponential smoothing method.Secondly,the sensitivity of the influencing factors to the displacement was calculated.The mean impact value was combined with the extreme learning machine with high predictive ability to analyze the changing relationship between the influencing factors and the displacement.The initial influencing factors were screened and the input quality was optimized.Grey relational analysis is used to calculate the correlation between the nodes of extreme learning machine,and the nodes with high correlation are selected to optimize the structure of grey extreme learning machine.Finally,the predicted trend term and periodic term were superimposed to obtain the predicted total displacement,and the Bootstrap method was used to predict the displacement interval,which demonstrated the effectiveness of the model and the predicted interval covered the fluctuation of displacement.(2)The traditional prediction model lacks deep feature extraction and the causal relationship between triggering factors and time series,and only considers a single landslide prediction model.This paper proposes a collaborative prediction model of support vector regression(SVR)and long-short-term memory network(LSTM)based on grey wolf optimizer(GWO).First,the cumulative displacement is decomposed into trend term and periodic term using empirical mode decomposition,and a polynomial function is used to fit the trend displacement components.Secondly,using empirical mode decomposition to decompose influencing factors into highfrequency sequences and low-frequency sequences,mining the deep information of features,selecting low-frequency sequences with periodic characteristics as candidate triggering factors,and using mutual information calculation and displacement correlation to select main triggering factors.Then,the SVR-LSTM collaborative prediction model is designed to improve the data fitting and prediction ability in the model,and realize the complementary advantages of the prediction model.Further,GWO is used to adaptive optimize the weights and biases in the SVR-LSTM model,and the GWO-SVR-LSTM collaborative optimization model is constructed to predict the periodic displacement.Finally,the predicted cumulative displacement is obtained by superimposing the trend term displacement and the periodic term displacement.Experiments show that the prediction accuracy of the model is high,which verifies that the model has excellent stability and adaptability.