Abstract: Dam deformation is the result of combined action of water pressure, temperature and other factors. The dam deformation monitoring data series is non-stationary and non-linear, and also correlated in time dimension. To fully mine the association of deformation monitoring data in long- and short-time span, the long- and short-term memory (LSTM) is proposed to predict dam deformation in this paper. To further improve the precision of prediction, the autoregressive integrated moving average (ARIMA) model is used to correct the residual sequence of LSTM prediction. The LSTM-ARIMA prediction model of dam deformation is thus established. With a concrete gravity dam as a case study, the predicted results of LSTM-ARIMA model is compared and analyzed against the predicted results of ARIMA model and Support Vector Machine (SVM). The result demonstrates that the LSTM-ARIMA model outperforms ARIMA and SVM, with the root mean squared error (RMSE) falling by 40.65% and 59.00%, respectively, and meanwhile the mean absolute error(MAE) by 35.49% and 55.60%, respectively, which implies that the LSTM-ARIMA model has high prediction accuracy.

Key words: prediction of dam deformation, long- and short-term memory(LSTM), ARIMA, LSTM-ARIMA combinatorial model, prediction accuracy