Deformation Forecasting of Pulp-Masonry Arch Dams via a Hybrid Model Based on CEEMDAN Considering the Lag of Influencing Factors

Abstract

Deformations in dam structures can have a critical impact on dam safety and life. Accurate methods for dam deformation prediction and safety evaluation are thus highly needed. Dam deformations can be predicted based on many factors. The analysis of these influences on the deformation of the dam reveals a problem that deserves further attention: dam deformation lags behind environmental factors of the water level and temperature as well as the time lag of the temporal dam deformation data. In this paper, a hybrid deep learning model is proposed to enhance the accuracy of dam deformation forecasting based on lag indices of these factors. In particular, dam deformations are predicted using deep networks based on gated recurrent units (GRUs), which can effectively capture the temporal characteristics of dam deformation. In addition, an improved particle swarm optimization (IPSO) algorithm is used for optimizing the GRU hyperparameters. Furthermore, the complete ensemble empirical mode decomposition with adaptive noise algorithm (CEEMDAN) and the partial autocorrelation function (PACF) are exploited to select the lag factor indices. The accuracy and effectiveness of the proposed CEEMDAN–PACF–IPSO–GRU hybrid model were evaluated and compared with those of other existing models in terms of four different evaluation indices (MAE, MSE, R2, and RMSE) and using 9-year historical data for the case of a pulp-masonry arch dam in China. The experimental results show that our model outperforms other models in terms of the deformation prediction accuracy (R2 increased by 0.16%–9.74%, while the other indices increased by 14.55% to reach 96.69%), and hence represents a promising framework for general analysis of dam deformations and other types of structural behavior.