A deep learning prediction model of DenseNet-LSTM for concrete gravity dam deformation based on feature selection
Ye Zhang,Wen Zhong,Yanlong Li,Lifeng Wen
DOI: https://doi.org/10.1016/j.engstruct.2023.116827
IF: 5.582
2023-09-20
Engineering Structures
Abstract:Dam deformation can comprehensively reflect the operational status of the dam. Thus, it is significant to build a dam deformation prediction model with high accuracy for the operation of concrete gravity dams. The concrete gravity dam is a complex and dynamic system. Based on the current prediction methods, such as the statistical models or machine learning (ML) models, it is a challenge to capture the complicated relationships between deformation and various features, as well as the features and time. As a result, we have proposed a deep learning model called DenseNet-LSTM, which combines the densely connected convolutional Network (DenseNet) and long short-term memory (LSTM) network. Meanwhile, correlation analysis and random forest (RF) are also adopted to evaluate and select deformation features. In the case study, the deformation monitoring data of multiple points at different elevations in the same section of a typical concrete gravity dam in southwestern China are selected to verify the model. For all the monitoring points, the correlation coefficient of the proposed model is more than 0.99. The result shows that the DenseNet-LSTM model can reveal the dynamic evolution process of deformation of the concrete gravity dam. Compared with the CNN-LSTM, LSTM, and ML-based models, the accuracy of the proposed model is higher and its generalization ability is better, which provides new methods for dam safety monitoring.
engineering, civil