A high-performance displacement prediction model of concrete dams integrating signal processing and multiple machine learning techniques

Highlights

• VMD can separate signals with different variation features from dam displacement.

• TSVR and GRU have outstanding nonlinear processing capabilities.

• The proposed decomposition-ensemble model has excellent predictive performance.

Abstract

Considering the strong nonstationarity and nonlinearity of the measured displacement, a displacement prediction model of concrete dams is proposed using variational mode decomposition (VMD), twin support vector regression (TSVR) and gated recurrent unit (GRU). Firstly, VMD is adopted to decompose the measured displacement into several signals according to the time series features, and sample entropy (SE) method is used to avoid over-decomposition. Secondly, K-means clustering algorithm is utilized to categorize the decomposed signals into trend, periodic, semi-periodic and stochastic signals, and the trend, periodic and semi-periodic signals are utilized to reconstruct the displacement. Thirdly, the reconstructed displacement is predicted with TSVR-based hydrostatic-seasonal-time (HST) model, and the stochastic signals are predicted with GRU network-based model. Considering the influence of parameters on modelling performance, seagull optimization algorithm (SOA) is employed to determine the optimal parameters of TSVR and GRU. Finally, a displacement prediction model is established by superimposing the predicted values of TSVR-based and GRU-based models. The measured displacement of a gravity dam is utilized to elaborate the implementation process and to test the validity of the proposed model. Results show that the model has excellent predictive performance, which demonstrates that the model is feasible and powerful for precisely predicting dam displacement.

Introduction

Concrete dams during the operation period are bound to suffer from various dynamic and static cyclic environmental loads, such as hydrostatic and temperature loads, and accidental disasters like earthquakes and catastrophic floods [1]. And such external forces would lead to the deterioration of material properties and the reduction of structural resistances. The operation safety of dams directly affects the benefit's exertion of the projects, national economy development and regional stability. Dam failure will pose a serious threat to the life security of citizens in the downstream [2,3]. Recently, with the exploitation of water resources and the development of dam construction technology, Xiaowan, Xiluodu, Jinping I and some other super high concrete dams have been put into operation. The project scale and difficulty in security management of such engineered systems have far exceeded those of dams in the previous generation. The operation safety issue of dams has aroused common concern of the whole society.

The measured data-driven structural health monitoring (SHM) is a crucial method to provide vital theoretical basis and technical support for safeguarding healthy operation of concrete dams. Displacement can intuitively reflect the comprehensive operation behavior of concrete dams under the synergistic action of dynamic changes of environmental loads and nonlinear evolutions of material and structural properties [4,5]. Therefore, the measured displacement-based SHM model has become a research hotspot for real-time diagnosis of health status, identification of structural damage and prediction of future operational behavior [6], [7], [8]. The measured displacement-based SHM model can be classified into HST and hydrostatic-thermal-time (HTT) models [9], [10], [11]. Among which, HST model usually employs multicycle harmonic functions to simulate the periodic variation behavior of thermal displacement. Due to the simple structure, HST model is popularly used for dam displacement prediction [7,10,11]. However, harmonic functions cannot reasonably depict the short-term dynamic fluctuation behavior of thermal displacement induced by the temperature difference. Therefore, HST model cannot well elaborate the detailed variation behavior of thermal displacement when the ambient temperature is obviously colder or warmer than usual [1,10].

The mathematical analysis method-based model established via the causal relationship between dam displacement and its explanatory variables is the most popular modelling approach of the measured displacement-based SHM model due to its high computational efficiency. And with the development of computer technology, a large number of machine learning (ML) techniques, such as support vector machine (SVM) [2,12], extreme learning machine (ELM) [4], radial basis function network (RBFN) [6], boosted regression trees (BRT) [8], long-short term memory (LSTM) [13,14] and TSVR [1], have been successfully utilized in the measured displacement-based SHM model nowadays. The ML techniques effectively overcome the problem of multicollinearity among the explanatory variables in the linear regression analysis methods, such as multiple linear regression (MLR) and stepwise regression (SR). And they perform with excellent predictive accuracy. The application of such ML techniques dramatically facilitates the development of the measured displacement-based SHM model. In addition, decomposition-ensemble models integrating multiple techniques further enriched the modelling approach of the measured displacement-based SHM models. And the decomposition-ensemble models exhibit superior predictive performance to the single technique-based models. For instance, Li et al. [5] proposed a distributed time series evolution-based displacement prediction model using locally estimated scatterplot smoothing, non-linear least squares and dynamic time warping. Zhang and Xu [15] adopted empirical mode decomposition (EMD) to decompose the measured displacement into several displacement components with different variation behaviors, and then the displacement components were predicted by the optimized SVM. Bian and Wu [16] utilized EMD to decompose the measured displacement into periodic and irreversible displacements, and tested the modelling performances of ELM, SVM and back propagation neural network on periodic displacement. Cao et al. [17] utilized VMD to decompose the measured displacement into several components, and the decomposed components were predicted with ELM.

Considering the strong nonstationary and nonlinear variation behavior of the measured displacement, and the insufficiency of HST model on the interpretative capability of thermal displacement, a high-performance displacement prediction model of concrete dams is proposed based on signal processing and multiple ML techniques. Firstly, VMD is adopted to decompose the measured displacement into several time series, and SE method is used to avoid over-decompostion. Secondly, according to the modelling factors in HST model, the decomposed signals are categorized into trend, periodic, semi-periodic and stochastic signals via K-means clustering algorithm. Meanwhile, the trend, periodic and semi-periodic signals are utilized to reconstruct the displacement. Thirdly, TSVR is adopted to establish HST-based prediction model for the reconstructed displacement, and the stochastic signals are predicted with GRU network-based model. To improve the modelling performance, the parameters of TSVR and GRU are optimized by SOA. Finally, a high-performance displacement prediction model is established by superimposing the predicted values of TSVR-based and GRU-based models. The modelling procedure of the proposed model is elaborated and the predictive accuracy is tested with the measured displacements of a gravity dam. Modelling results demonstrate that the proposed model has excellent predictive performance. A novel prediction technique is provided for accurately predicting dam displacement.