关键词：
Deep learning
摘要：
This study proposes a refined prediction method that leverages mutual feedback of spatiotemporal modal information from multiple monitoring points. This approach addresses the limitations of traditional step-like landslide displacement prediction methods regarding granularity and timeliness at annual and monthly scales, as well as the insufficient accuracy arising from increased nonlinearity and redundant information interference at the daily scale. Initially, the variational modal decomposition method, optimized using the sparrow search algorithm (SSA-VMD), is employed to minimize the system’s redundant information entropy. The displacement sequence is decomposed into trend term displacements that reflect material properties, along with periodic and random term displacements influenced by environmental factors. Subsequently, Pearson correlation analysis and Granger causality analysis are performed to identify significant statistical associations between the displacement modes and ecological factors, as well as to filter relevant statistical relationships between displacement modes and environmental variables. This is followed by the design of a CNN-Informer deep learning structure. Spatiotemporal information is extracted using convolutional neural networks (CNN), while the sparse attention mechanism and self-attention distillation mechanism of the informer are utilized to mitigate interference from redundant information and dynamically capture long-term dependency relationships between features and displacements. The ultimate outcome is a point-level daily displacement forecast achieved through time-series reconstruction. This study utilizes daily monitoring data from the monitoring points GSCX3, ZGX111, and GSCX5, collected from October 2017 to December 2021, to analyze the step-like Bazimen landslide. The results indicate that: (1) The model achieves RMSE values of 2.61, 3.84, and 3.91 mm for the three monitoring points, demonstrating a 71.94% reduction in RMSE a
欢迎来到三峡大学图书馆！
中文科技期刊
详细信息
