关键词: 大坝变形, 组合预测模型, 集合经验模态分解, 长短期记忆神经网络, 多元线性回归

Abstract: Under the combined action of water pressure, temperature and aging, dam deformation monitoring data series is non-stationary and nonlinear. We introduced the method of Ensemble Empirical Mode Decomposition (EEMD) to process deformation data and obtained the multi-scale dam deformation components. For the complex high frequency components, we employed the Long and Short Term Memory network (LSTM) to achieve better prediction results; for low frequency components with periodic changes, we adopted Multiple Linear Regression (MLR) for rapid and effective prediction. Through the refactoring of components we can acquire the predicted result of dam deformation. Analysis of engineering examples demonstrated that the EEMD method avoided modal aliasing and attained more reasonable multi-scale deformation components. LSTM and MLR have good predictive ability for high and low frequency components respectively. By comparing the result of components superposition separately with those of a variety of single prediction algorithms, EMD decomposition algorithm, and traditional models, we found that the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and the Root Mean Square Error (RMSE) of the EEMD-LSTM-MLR combinatorial model were lower than the comparative models above, indicating higher prediction precision.

Key words: dam deformation, combinatorial prediction model, Ensemble Empirical Mode Decomposition, Long and Short Term Memory network, Multiple Linear Regression