An end-to-end deep learning framework with nonstationary weighting strategy for enhancing reservoir landslide displacement prediction

Accurate prediction of temporal reservoir landslide displacement using key environmental variables, such as rainfall depth, reservoir water level, and groundwater level, is crucial for early warning. These variables exhibit distinct, time-varying influences on displacement, yet conventional deep learning models often overlook such dynamics by adopting stationary sliding windows and assigning equal importance to all variables and timesteps, thereby limiting forecasting accuracy. To address this limitation, we propose a novel Dynamic Weight Generator (DWG) that learns per-variable, per-timestep weights within a user-defined historical window. DWG generates a weight map applied element-wise to normalized inputs before they are fed into a prediction backbone, enabling the model to emphasize critical variables and influential timesteps while down-weighting less relevant ones. DWG is integrated into representative deep-learning models, including convolutional neural networks, long short-term memory networks, transformers, and Mamba architectures. All models are trained end-to-end, allowing the weight map to be optimized directly for displacement prediction. Compared with conventional stationary approaches, the proposed framework consistently improves accuracy across all backbones. Overall, DWG effectively captures dynamic, variable-dependent lag effects, enhancing temporal landslide displacement prediction.