Abstract: To handle the massive dam safety monitoring data, quick and reasonable determination of the variable factors of a dam deformation prediction model can effectively improve prediction efficiency and accuracy. This paper constructs a dam deformation prediction model by combining the least absolute shrinkage and selection operation (LASSO) variable selection and a long short-term memory (LSTM) network. First, a set of relevant influencing factors of dam deformation is determined through analysis of the dam deformation mechanism; then, the LASSO algorithm is used to remove the insignificant factors and select the optimal influencing factors as the model input variables, and a dam deformation prediction model is constructed using the LSTM network. Finally, this new method is verified and discussed with application to a case study of a roller compacted concrete gravity dam of Zaoshi Water Control. The results show it improves the accuracy significantly with relatively small mean absolute errors (MAE), mean square errors (MSE) and root mean square errors (RMSE). Compared with the conventional model, its variable selection based on the LASSO algorithm makes the model construction simpler and more efficient, and thus it is conducive to processing and analysis of massive dam monitoring data.

Key words: dam deformation, variable selection, least absolute shrinkage and selection operation algorithm, long short-term memory, prediction model