PCA-IEM-DARNN: An enhanced dual-stage deep learning prediction model for concrete dam deformation based on feature decomposition

Abstract

Deformation is a crucial indicator of structural integrity, essential for ensuring the long-term safety of dams. Existing models face challenges in accurately simulating the strong fluctuation and non-stationary characteristics of concrete dam deformations, along with in effectively extracting useful information from input data. Therefore, this study proposes the PCA-IEM-DARNN model, integrating feature decomposition and dual attention mechanisms for deformation prediction. Initially, the ICEEMDAN method decomposes displacements into IMFs, facilitating deep analysis of deformation sequence fluctuations. These IMFs are then categorized into reconstructed displacements and stochastic signals, simplifying the complexity of the sequence. The DARNN model is employed to predict these components separately, extracting critical information from input data. Featuring dual attention mechanisms, the model adapts to select relevant input features and capture time series dependencies, significantly enhancing prediction accuracy. Experimental results show an average R2 of 0.991, surpassing comparative models and offering valuable insights for future concrete dam deformation predictions.