Research and analysis of the TCN-Multihead-Attention prediction model of landslide deformation in the Three Gorges Reservoir area, China

Huan Chen¹Yimin Liu^1*Yixuan Li²

• ¹School of Intelligent manufacturing, Chengdu Technological University, Chengdu, China
• ²Japan Advanced Institute of Science and Technology, Nomi, Ishikawa, Japan

Landslide deformation prediction is a crucial task in geotechnical engineering and disaster prevention. Developing an accurate and reliable landslide displacement prediction model is vital for effective landslide warning systems. This paper proposes a TCN-Multihead-Attention prediction model for landslide deformation based on temporal convolutional networks (TCNs). We collected eight years of monitoring data from the Huangniba Dengkan landslide in the Three Gorges Reservoir area, including surface deformation (horizontal displacement and elevation), rainfall, and reservoir levels. A comprehensive analysis was conducted to assess the effects of rainfall, reservoir levels, and elevation on landslide horizontal displacement. Utilizing the multi-input and single-output characteristics of the long-period time series dataset, we developed the TCN-Multihead-Attention prediction model of landslide deformation. Model evaluation demonstrated that the coefficient of determination (R²) for the test set reached 0.995, with MAPE and RMSE at only 0.482 and 7.180, respectively, indicating high accuracy. Additionally, we developed other prediction models based on single TCN, Attention-based Transformer, RNN-based LSTM, and the hybrid CNN-BiLSTM for comparison. Compared with existing models, the TCN-Multihead-Attention model integrates dilated causal convolutions from TCN with multi-head attention to effectively fuse nonlinear interactions of multi-source environmental factors, capture long-term evolutionary trends, and accurately identify local mutation patterns, demonstrating superior reliability for landslide deformation forecasting in reservoir regions.