基于贝叶斯优化LSTM‑Transformer模型的热带台风暴雨区库岸边坡变形预测OA

A hybrid deep learning prediction model integrating Bayesian optimization(BO)with LSTM-Transformer is proposed for reservoir bank slope deformation prediction.The LSTM network is employed to hierarchically extract local temporal variation features associated with rainfall infiltration and historical deformation,while the multi-head self-attention mechanism of the Transformer is used to capture global dependencies of the deformation field across spatiotemporal scales.Bayesian optimization is introduced to achieve adaptive hyperparameter tuning,and a Pearson correlation coefficient matrix is constructed to perform weighted initialization of the attention mechanism,thereby enhancing the model's ability to focus on key monitoring points with coordinated deformation behavior.The proposed model is validated using a water-related slope of a hydraulic engineering project in Hainan Province as a case study.The experimental results show that,compared with standalone LSTM and Transformer models,the BO-LSTM-Transformer model effectively suppresses error accumulation in long-sequence predictions and significantly improves prediction accuracy during deformation stages.This paper provides a high-precision and interpretable prediction tool for intelligent monitoring of reservoir bank slopes under extreme rainfall conditions.