Landslide Displacement Prediction Model Based on Time Series and CNN-GRU

Abstract

Landslide displacement prediction is an important basis for early landslide warning.This paper proposes a prediction model of landslide moving states based on time series and convolutional gated recurrent unit (CNN-GRU) to deal with the shortcomings of previous prediction models.Firstly,after employing wavelet analysis to determine the displacement of the trend term,the exponential smoothing method is adopted to decompose the cumulative displacement to obtain two displacement types of the trend term and the periodic term,and the trend term is fitted by a five-order polynomial.Then,the autocorrelation function is utilized to test the periodic displacement characteristics,and the gray correlation method is applied to determine the correlation degree between each factor and the periodic term.Meanwhile,the periodic term and the influencing factor are input into the CNN-GRU model for prediction,and finally the predicted cumulative displacement value is obtained by superposition.By taking the Baishui River landslide in the Three Gorges Reservoir area as an example,this paper selects the data from January 2004 to December 2012 for study,and the average absolute error percentage of the final prediction results is only 0.525%,with RMSE of 9.614 and R2 of 0.993.Experimental results show that CNN-GRU has higher prediction accuracy.

Keywords

• displacement prediction;time series;convolutional gated recurrent unit;Baishui River landslide