Interpretable landslide displacement prediction based on KAN-N-Beats

Aiming at problems of low prediction accuracy, poor generalization, high model complexity in existing landslide displacement prediction models, and poor interpretability of traditional deep learning algorithms, this study proposes a landslide displacement prediction model based on KAN-N-Beats (KAN-N-Beats).KAN-N-Beats replaces the fully connected layers of the N-Beats architecture with Kolmogorov-Arnold Networks (KAN).Using KAN′s adaptive learning capabilities enhances both prediction accuracy and generalization performance.Furthermore, KAN intrinsically improves model interpretability through techniques like sparsity regularization, visualization, pruning, symbolic regression, and affine function fitting.While, N-Beats decomposes the prediction task into interpretable trend and seasonal components, which enables extraction of distinct time-series features.This synergistic combination endows KAN-N-Beats with significantly higher interpretability of its predictions.Additionally, N-Beats′ inherent decomposition capability allows for predicting landslide displacement after decomposition, eliminating the need for extensive feature engineering, which reduces model complexity and improves prediction efficiency.Data from the Baishuihe landslide and Bazimen landslide in the Three Gorges Reservoir area (sourced from the National Cryosphere Desert Science Data Center) served as the research dataset.Validation results at monitoring point ZG118 of the Baishuihe landslide are highly consistent with the real displacement, showing R² and RMSE values of 0.988 7 and 5.031 3, respectively.The generalization test at the monitoring points of Baishuihe landslide ZG118 and Bazimen landslide ZG110 and ZG111 shows that the proposed algorithm is better than other compared models, which can improve the accuracy of landslide prediction and has interpretability.The research results can provide a reference for improving the efficiency of landslide displacement prediction.