Deformation prediction of arch dams by coupling STL decomposition and LSTM neural network

Deformation plays an important role in dam monitoring, especially in judging potential damage. Excessive deformation serves as an indicator of the abnormal evolution of dam structures, but the pivotal issue lies in definition for “excessive” precisely. Deformation prediction serves as reference for detection, but it is challenged by multiple factors. Sequence decomposition is an approach to select related factors and predict precisely, but the mechanism in application is a critical problem. This study proposes a decomposition Long short-term memory (DLSTM) model for the deformation prediction of arch dams. Segmented correlation feature extraction method balances feature correlation and consistency. Multiple linear regression is introduced for the integration, and the periodic extension is used to simplify unnecessary predicting operations. The proposed DLSTM model is evaluated using monitoring data from the LYX arch dam. The results show an 18% improvement in the accuracy, as well as better stability of the prediction model. The DLSTM-based multiple linear regression outperforms nonlinear neural network methods in terms of both physical interpretation and performance. The superiority of the proposed model over BP neural network, wavelet neural network, and LSTM model demonstrates that the DLSTM model is a promising approach for practical applications in dam deformation monitoring.