A dual-stage combined displacement prediction model for concrete dam based on adaptive time series decomposition noise reduction and residual chaotic feature separation

Abstract

High-precision analysis and prediction of dam displacement is a crucial strategy to grasp the working attitude of dams and diagnose dam anomalies. However, the existing models often fail to accurately identify the interference noise existing in the form of short-frequency and small-fluctuations, resulting in the masking of the true deformation features. Meanwhile, existing studies often focus on one-stage prediction models, discarding the rich and valuable information contained in the residual sequence. Furthermore, the existing dual-stage models often fail to deeply consider the chaotic characteristics existing in the residuals. Therefore, this paper proposes a dual-stage combined displacement prediction model for concrete dam identifying the displacement sequence interference noise and considering the chaotic characteristics of the residual sequence. Firstly, the adaptive noise complete empirical mode decomposition, the improved sparrow search algorithm and the threshold evaluation index are combined to adaptively achieve the optimal decomposition noise reduction and retain the effective deformation features. Secondly, a gradient boosting tree is utilized to fit the effective component and combine it with the processed noise component to build a high-quality residual sequence that is rich in information. Thirdly, the residual sequence is decomposed into intrinsic mode functions with different temporal characteristics by utilizing the optimized variational mode decomposition. Finally, construct a chaotic time series based on chaos theory. Taking the sample entropy as the basis of judgement, for high-frequency components, the gradient boosting tree algorithm is utilized to capture their dynamic features. For low-frequency components, the depth-separable convolutional neural network, multi-head attention mechanism and bidirectional long short-term memory neural network are organically combined to comprehensively learn the deformation features. Case analysis shows that the RMSE of the model proposed in this paper in the measurement point PL11–5 sequence has reached an astonishing 0.0794, and the maximum improvement degree compared with the control model has reached 79.35 %. The results show that this model obtains strong generalization ability and high robustness, and can provide reference for dam safety monitoring.