| Landslides are one of the most common natural geological disasters,characterized by strong destructive force,high frequency,and fast movement,and accompanied by disasters such as floods and mudflows,seriously causing massive economic losses,damaging the public environment,and threatening people’s lives.Various triggering factors influence landslide displacement and best reflect the evolution process.When the landslide displacement reaches a certain value,the landslide will erupt.Therefore,the study of landslide displacement prediction has become a hot spot in engineering.In this paper,Long Short-Term Memory(LSTM)and Generative Adversarial Nets(GAN)are used in combination with Variational Mode Decomposition(VMD)algorithm to construct prediction models.The details are as follows.(1)Since landslide displacements are influenced by various induced factors,which makes the data distribution show a step type with nonlinear and non-smooth characteristics,such irregularly varying displacement series make it challenging to fit the prediction model.To make the input data have certain regular variations,this study adopts the VMD algorithm to preprocess the landslide displacement time series and decompose them into trend term displacement,periodic term displacement,and random term displacement.The original data are transformed from the time domain to the frequency domain so that the decomposed subseries have more practical physical meaning and effectively combine the monitoring point environment with the displacement evolution trend.At the same time,it makes the distribution characteristics of each displacement curve after decomposition obvious.It is convenient for the prediction model to predict each subsequence in the subsequent experiments and get more accurate prediction values.(2)To address the limitations of a single neural network model in step-type landslide displacement prediction,it is difficult to fit the dynamic evolution process of landslide displacement and other problems.To effectively learn the information of landslide displacement changes in a long period,this paper proposes a new combined VMD-LSTM model for single-step prediction of landslide displacement,in which the LSTM network has the functions of memory and filtering information and high efficiency.The VMD decomposes the landslide displacement data into trend term displacement,periodic term displacement,and random term displacement,which makes the LSTM model extracts more internal features in different frequency domains,improving the model’s prediction accuracy.Taking the landslide of the Baishui River in the Three Gorges reservoir area as an example,the prediction error of the VMD-LSTM model is reduced by 9.75 compared with the traditional LSTM model,which has higher prediction accuracy.(3)To further expand the types of landslide displacement prediction models,explore the causality of time series information,and address the problems such as error accumulation in landslide displacement prediction,this paper proposes a new VMD-GAN model for landslide displacement time series prediction.The traditional GAN model can deal with the visual domain well,but it still needs to be improved in the time series domain.At the same time,the training process is unstable,which can easily cause the model to crash.Therefore,the structure of the GAN model is improved,the generator is built by Temporal Convolutional Network(TCN),and the discriminator is built by Convolutional Neural Networks(CNN).Through the adversarial training between them,the GAN is effective.The problem of the classical GAN model is solved by adversarial training between the two,which effectively extends GAN to time series.Taking the landslide of the Baishui River in the Three Gorges reservoir area as an example,compared with the VMD-LSTM model proposed in the previous section,and the traditional LSTM,Gate Recurrent Unit(GRU),Back Propagation Neural Network(BPNN)and TCN,VMD-GAN has the best prediction accuracy and fitting ability.In conclusion,the method proposed in this paper can provide some early warning for solving the problem of landslide disasters. |
