Abstract: Landslide disasters can be warned based on monitoring and prediction of displacements. In view of the complex internal mechanisms of landslides, data-driven model is an effective approach of simulating landslide evolvement when the precise models reflecting the mechanisms cannot be obtained. Considering the complex nonlinear dynamics of landslides, we built a recurrent dynamic neural network for landslide displacement based on reservoir computing. Furthermore, we further employed fractal interpolation to enhance the reservoir training process and expand the displacement data sets. The method was used to predict the developments of three different typical landslides, and the predictions are all very close to the actual measurements. It is a good solution for complex dynamic prediction with short-time sequence.

Key words: landslide, reservoir computing, recurrent neural network, fractal interpolation